betaFeedback

WorksComputer Science

Entropy-Regularized Integrated Gradients for Stable Feature Attribution

Abstract

Path-based attribution methods like Integrated Gradients (IG) are widely used to interpret neural networks, yet imperceptible input perturbations can drastically alter their attribution maps while leaving model predictions unchanged. We introduce Entropy-Regularized Integrated Gradients (ERIG), a path-based attribution method that intrinsically attenuates gradient irregularities along the integration path using maximum-entropy softmax weighting with temperature parameter ττ. We quantify the effect of the entropy-regularized weighting on attribution stability by using the Lipschitz properties of the model and its gradients. We decompose the attribution change under an input perturbation into a model-response term, whose bound is ττ-independent, and a weight-variation term, whose bound decreases at rate O(1/(2τ))\mathcal{O}(1/(2τ)). Their ratio R(τ)R(τ) characterizes the trade-off between gradient information and attribution stability. We further add a new dimension to attribution evaluation by adapting classical noise estimators into two metrics, LINV-C and LINV-P, that quantify attribution noise. Across eleven CNN and Transformer backbones on ImageNet and Oxford-IIIT Pet datasets, our experiments show that, relative to the next-best method, ERIG remains comparable on ranking-based fidelity metrics while improving localization by approximately 24% and attribution stability by approximately 47-fold on average, reaching up to four orders of magnitude improvement over IG, all at negligible additional computational cost.

Introduction

Deep learning has advanced a myriad of fields such as vision [80, 86], language [34], science [45, 63, 64], and archaeology [60]. Yet these models remain black boxes, offering little insight into their predictions. High-stakes applications demand transparent explanations and raise concerns about model biases [15, 54] and fairness [83].

Explainable Artificial Intelligence (XAI) has emerged to address this lack of transparency [2, 30], yet existing methods carry critical weaknesses [50]; Ghorbani et al. [28] demonstrate that imperceptible input perturbations can drastically alter gradient-based explanations while leaving predictions unchanged. The root cause is the irregular input-space gradient field ∇f\nabla f of neural networks. Standard deep networks exhibit excess input-space curvature [20, 49, 71], so minuscule perturbations produce large gradient shifts [29, 73], and the gradient field carries high-frequency components misaligned with semantic structure [24, 77, 95]. This irregularity is baked in by training. Standard objectives do not regularize input gradients; gradients shatter toward white noise as depth grows [7], sparse activation patterns concentrate gradient information in a few noise-sensitive directions [32], and while batch normalization smooths the optimization landscape [61], it does not constrain input gradients explicitly. In this context, path integration averages out some of these fluctuations, but the effect is incomplete, and attribution maps still inherit substantial noise [3, 5, 8, 28, 44, 94].

For model predictions, diverse ensembles can improve accuracy [90] and robustness [84, 91]. At the explanation level, SmoothGrad [68] addresses these issues post hoc by averaging gradients over local input perturbations to improve stability and reduce noise. This motivates explicitly regularizing gradient noise for reliable explanations.

We design Entropy-Regularized Integrated Gradients (ERIG), a path-based feature attribution method that addresses the question: How can we weight gradient contributions along the integration path so that attribution noise is suppressed while genuine model sensitivity is preserved? Intuitively, ERIG answers this by not passing raw gradients directly into the attribution. Instead, it converts them into weights that stay close to uniform unless the local sensitivity warrants concentration. Here, a single temperature τ\tau controls how readily the weights concentrate. We provide a summary of ERIG in Figure 1. Unlike IG [72] and IDGI [93], which couple weights directly to raw gradients, ERIG uses a maximum-entropy softmax over squared gradients as an indirect sensitivity weighting. A formal Lipschitz analysis then reveals that τ\tau selectively controls how gradient perturbations propagate through to the weights. The associated ratio R(τ)R(\tau), defined in § 3.2, quantifies this selectiveness. Our contributions are threefold:

Entropy-Regularized Integrated Gradients (ERIG) framework

Figure 1. Entropy-Regularized Integrated Gradients (ERIG). An integration path Γ(t)\Gamma(t) connects the baseline xbx^b to the input xax^a. At each path point x(k)x^{(k)}, the indirect weighting scheme maps the raw gradient g=∇fg = \nabla f to importance weights wi∗w_i^* through a softmax over the squared gradients g⊙gg \odot g with temperature τ\tau. Intuitively, the weights stay close to uniform unless the local sensitivity warrants concentration, and τ\tau controls how readily they concentrate. Summing the weighted contributions wi∗⋅δf(k)w_i^* \cdot\delta f^{(k)} over the path yields the final attribution ϕi\phi_i.

  • We define an indirect weighting scheme and leverage a maximum-entropy principle to derive a closed-form softmax expression with a temperature parameter τ\tau that controls the method’s sensitivity to gradient irregularities.

  • We prove a decomposition theorem establishing that τ\tau attenuates the weight-variation component of attribution perturbation at rate O(1/(2τ))O(1/(2\tau)), while the model-response component admits a τ\tau-independent upper bound.

  • We adapt established noise estimators [16, 39] to the attribution setting, yielding two attribution noise metrics, LINV-C and LINV-P. This adds a new noise dimension to traditional quantitative attribution evaluations.

Related Work

Since its introduction, Integrated Gradients (IG) [72] has inspired a broad family of extensions. Some improve its numerical approximation [66], while others adapt it to temporal data [40], NLP [22], and learned manifolds [96]. Further works decompose it at the neuron or layer level [19], combine it with segmentation [42], or connect it to game theory [17, 79]. Another line of work modifies the baseline or the path. For example, Blur-IG [92] blurs the input progressively, Guided IG [43] selects low-gradient path points, and Expected Gradients [23] averages over data-sampled baselines. Iterative counterfactual updates [97] and geometric optimization [65] optimize the baseline, while Shapley-inspired sampling [47, 74] eliminates baseline dependence. Closest to our work, IDGI [93] weights gradient contributions by gi2g_i^2, demonstrating that squared-gradient weighting suppresses attribution noise relative to IG.

A complementary line of work tailors attribution to specific architectures. AttnLRP [1] derives relevance propagation rules for Transformer attention, OMENN [87] represents piecewise-linear networks as a single input-dependent matrix, DAVE [88] scores ViT attributions through a distribution-aware gradient decomposition, and LibraGrad [48] balances Transformer gradient flow to improve all gradient-based attributions. Such modified backward computations are known to require careful interpretation [67]. ERIG is instead architecture-agnostic; it uses the exact autograd gradients of any differentiable model, so its theoretical guarantees hold for the deployed model as is. LibraGrad in particular composes with ERIG, correcting how gradients are computed while ERIG weights these corrected gradients. We leave this combination to future work.

Background

A General Framework

Let f:RN→Rf : \mathbb{R}^{N} \to\mathbb{R} be a neural network, and let Γ:[0,1]→RN\Gamma: [0,1] \to\mathbb{R}^{N} be an integration path from baseline xb=Γ(0)x^b = \Gamma(0) to input xa=Γ(1)x^a = \Gamma(1). At each path point tt, the gradient vector is g(t)=∇f(Γ(t))g(t) = \nabla f(\Gamma(t)), with components gi(t)=∂f∂xi∣Γ(t)g_i(t) = \frac{\partial f}{\partial x_i}\big|_{\Gamma(t)}. For a weighting function w:RN→ΔN−1w : \mathbb{R}^{N} \to\Delta^{N-1} mapping the gradient vector to the NN-dimensional probability simplex, the attribution for feature ii, inspired by path-based methods, is

ϕi=∫Γwi(g(t)) df(t)≈∑k=1Kwi(g(x(k))) δf(k),(1)\phi_i = \int_{\Gamma} w_i(g(t))\,\mathrm{d}f(t) \approx\sum_{k=1}^{K} w_i(g(x^{(k)}))\,\delta f^{(k)}, \tag*{(1)}

where {x(k)}k=0K\{x^{(k)}\}_{k=0}^{K} discretizes Γ\Gamma and δf(k)=f(x(k))−f(x(k−1))\delta f^{(k)} = f(x^{(k)}) - f(x^{(k-1)}). Different choices of ww yield different methods; Integrated Gradients (IG) [72] approximately corresponds to wi∝giw_i \propto g_i and Important Direction Gradient Integration (IDGI) [93] to wi∝gi2w_i \propto g_i^2. Here one can ask: what is the optimal choice of ww?

Indirect Gradient Weighting

Both IG and IDGI couple attribution weights directly to the gradient, leaving no controllable mechanism to attenuate irregularities or noise in the gradient field. Guided IG [43] takes the opposite approach, reducing attribution noise by adaptively steering the integration path away from high-gradient regions; however, this discards potentially informative gradient signal along with the noise. A key observation is that introducing a weighting function w(g)w(\mathbf{g}) that maps the gradient vector to the simplex adds a layer of indirectness between the raw gradient field and the final attribution. This creates a design space that controls how much gradient sensitivity propagates into the attributions. In our specific design, a single temperature τ\tau governs this trade-off.

Entropy-Regularized Integrated Gradients

Maximum-Entropy Sensitivity Weighting

Given the indirect weighting framework established in § 2, we seek a weighting scheme ww that, for every point on the integration path, is maximally unbiased (maximizes entropy [41]) while still concentrating weight on locally sensitive features. Formally, we impose the constraint ∑i=1Nwigi2≥γ\sum_{i=1}^{N} w_i g_i^2 \geq\gamma, keeping the expected local sensitivity under ww above a threshold. The choice of gi2g_i^2 follows derivative-based global sensitivity measures [70]. Locally, a first-order Taylor expansion shows that a small unbiased perturbation of feature ii induces output variance proportional to (∂f/∂xi)2(\partial f/\partial x_i)^2 (Appendix A.§ 1), making gi2g_i^2 a natural local sensitivity measure. We then have,

Proposition 1 (Maximum-entropy Weights). For the pointwise entropy-maximization problem at any point on Γ(t)\Gamma(t),

w∗=arg⁡max⁡w∈ΔN−1(−∑i=1Nwiln⁡wi)s.t.∑i=1Nwigi2≥γ.(2)\mathbf{w}^{*} = \underset{\mathbf{w}\in\Delta^{N-1}}{\arg\max}\left(-\sum_{i=1}^{N}w_i\ln w_i\right)\quad\text{s.t.}\quad\sum_{i=1}^{N}w_i g_i^2 \geq\gamma. \tag*{(2)}

The unique optimum solution is

wi∗=exp⁡(gi2/τ)∑j=1Nexp⁡(gj2/τ)=Softmax⁡i(∇f⊙∇fτ),(3)w_i^{*}=\frac{\exp(g_i^2/\tau)}{\sum_{j=1}^{N}\exp(g_j^2/\tau)}=\operatorname{Softmax}_i\left(\frac{\nabla f\odot\nabla f}{\tau}\right), \tag*{(3)}

where τ>0\tau>0 satisfies ∑i=1Nwi∗gi2=γ\sum_{i=1}^{N}w_i^{*}g_i^2=\gamma.

Proof. See Appendix A.§ 1.1.

In Eq. (3), ⊙\odot denotes the elementwise (Hadamard) product, and the subscript in Softmax⁡i\operatorname{Softmax}_i selects the ii-th component of the softmax output. Then ERIG attribution for feature ii is,

ERIG⁡i(Γ;τ)=∫ΓSoftmax⁡i(∇f⊙∇fτ) df.(4)\operatorname{ERIG}_i(\Gamma;\tau)=\int_{\Gamma}\operatorname{Softmax}_i\left(\frac{\nabla f\odot\nabla f}{\tau}\right)\,\mathrm{d}f. \tag*{(4)}

In practice, τ\tau controls the concentration of the weights. Smaller τ\tau focuses attribution on the few features with the largest gradients; larger τ\tau spreads the attribution more uniformly losing discriminability.1 While Proposition 1 is a

per-path-point optimization, making τ\tau in principle a per-point parameter τ(t)\tau(t), we make the simplifying choice of a constant τ\tau; Appendix A.12 ablates a path-dependent schedule for comparison. See Appendix A.9 for ERIG algorithm.

Intrinsic Properties. The max-entropy formulation endows ERIG with attribution stability and improved localization, and Figure 2 makes its two mechanisms directly observable by tracking each method’s weight concentration along the integration path. The movement of the model output splits the path into an initial flat segment, a region of rapid output change, and a saturation plateau, and ERIG is the only method whose concentration follows this structure. Over the flat and saturated segments, where gradient spikes are noisier, entropy regularization [14, 56] keeps its weights closest to uniform, suppressing the noise (NS) that purely gradient-proportional schemes propagate unchecked. Where the output changes rapidly and the class evidence lies, the sensitivity constraint γ\gamma concentrates them to IDGI’s level, enhancing the signal (SE). The resulting attributions (Figure 3) confirm this qualitatively, with noise visibly suppressed and discriminative features highlighted. Further, § 4.7 validates them quantitatively via ablations. These properties are intrinsic to the method and add no model evaluations beyond the KK path points; ensemble methods like SmoothGrad [68] are instead extrinsic, needing S×KS\times K evaluations, where SS is the number of noise samples.

Figure 2. Weight concentration along the integration path (100 ImageNet validation images, K=200 steps). Top, mean output fˉk\bar{f}_k and increment magnitude ∣δfk∣|\delta f_k|; bottom, mean Gini coefficient of the per-step weight (w(g))(w(\mathbf{g})) maps. Only ERIG (τ=3\tau=3) suppresses noise over flat and saturated segments (NS) and enhances signal over the informative segment (SE).

MobileNetV4 visualizations

Figure 3. MobileNetV4 Visualizations — More examples in A.3

Decomposition of Attribution Perturbation

We now formalize how gradient perturbations propagate into the ERIG attribution. The key insight is that the attribution change under an input perturbation decomposes into two components, only one of whose bounds depends on τ\tau. We adopt two standard assumptions,

Assumption 1 (Lipschitz model). ff is LfL_f-Lipschitz: ∣f(x′)−f(x)∣≤Lf∥x′−x∥2|f(x')-f(x)| \le L_f\|x'-x\|_2 for all x,x′∈RNx,x' \in\mathbb{R}^N.

Assumption 2 (Lipschitz gradients). ∇f\nabla f is LgL_g-Lipschitz: ∥∇f(x′)−∇f(x)∥2≤Lg∥x′−x∥2\|\nabla f(x')-\nabla f(x)\|_2 \le L_g\|x'-x\|_2 for all x,x′∈RNx,x' \in\mathbb{R}^N.2

Consider a straight-line path Γ(t)=xb+t(xa−xb)\Gamma(t)=x^b+t(x^a-x^b) for t∈[0,1]t\in[0,1]. Under perturbation xa→x~a=xa+ηx^a\to\tilde{x}^a=x^a+\eta, the perturbed path is Γ~(t)=Γ(t)+tη\tilde{\Gamma}(t)=\Gamma(t)+t\eta, and the attribution change, setting aside O(∥η∥22)\mathcal{O}(\|\eta\|_2^2) terms for a small enough ∥η∥2≪1\|\eta\|_2\ll1, decomposes via the product rule as

ΔERIGi≈∫01[wi∗(g(t);τ) Δf˙(t)⏟model-response+Δwi∗(g(t);τ) f˙(t)⏟weight-variation] dt(5)\Delta\mathrm{ERIG}_i \approx\int_0^1 \left[\begin{aligned} &\underbrace{w_i^*(\boldsymbol{g}(t);\tau)\,\Delta\dot{f}(t)}_{\text{model-response}}\\ &\quad+\underbrace{\Delta w_i^*(\boldsymbol{g}(t);\tau)\,\dot{f}(t)}_{\text{weight-variation}} \end{aligned}\right]\,dt \tag*{(5)}

where Δwi∗=wi∗(g(Γ~(t));τ)−wi∗(g(Γ(t));τ)\Delta w_i^*=w_i^*(\boldsymbol{g}(\tilde{\Gamma}(t));\tau)-w_i^*(\boldsymbol{g}(\Gamma(t));\tau), and f˙(t)=df(Γ(t))/dt\dot{f}(t)=df(\Gamma(t))/dt is the function increment rate along the path, with Δf˙(t)=f˙(Γ~(t))−f˙(Γ(t))\Delta\dot{f}(t)=\dot{f}(\tilde{\Gamma}(t))-\dot{f}(\Gamma(t)) as its change under the perturbation. Please note that df=f˙ dtdf=\dot{f}\,dt in Eq. (4).

The model-response term captures the change in function increment rate along the perturbed path. The weight-variation term captures how gradient changes induced by the input perturbation propagate through the weighting scheme, redistributing the original function increments across features. Let M,W∈RNM,W\in\mathbb{R}^N denote the model-response and weight-variation vectors with components Mi=∫01wi∗Δf˙ dtM_i=\int_0^1w_i^*\Delta\dot{f}\,dt and Wi=∫01Δwi∗f˙ dtW_i=\int_0^1\Delta w_i^*\dot{f}\,dt, so that ΔERIG≈M+W\Delta\mathrm{ERIG}\approx M+W. Theorem 2 bounds these components:

Theorem 2 (Selective Attenuation). Let ∣Γ∣=∥xa−xb∥2|\Gamma|=\|x^a-x^b\|_2 denote the path length. Under Assumptions 1 and 2,

∥ΔERIG∥2≤∥M∥2+∥W∥2≤(Lf+12Lg∣Γ∣)∥η∥2⏟bound on ∥M∥2⋅(1+R(τ)),(6)\begin{aligned} \|\Delta\mathrm{ERIG}\|_2 &\le\|M\|_2+\|W\|_2\\ &\le\underbrace{\left(L_f+\tfrac{1}{2}L_g|\Gamma|\right)\|\eta\|_2}_{\text{bound on }\|M\|_2}\cdot\left(1+R(\tau)\right), \tag*{(6)} \end{aligned}

where

R(τ)=Lf2Lg∣Γ∣2τ(Lf+12Lg∣Γ∣)+O(∥η∥2).(7)R(\tau)=\frac{L_f^2L_g|\Gamma|}{2\tau\left(L_f+\tfrac{1}{2}L_g|\Gamma|\right)}+\mathcal{O}(\|\eta\|_2). \tag*{(7)}

Here Lf2Lg∣Γ∣∥η∥2/(2τ)+O(∥η∥22)L_f^2L_g|\Gamma|\|\eta\|_2/(2\tau)+\mathcal{O}(\|\eta\|_2^2) is the bound on the weight-variation term ∥W∥2\|W\|_2.

Proof. See Appendix A.7.

The model-response bound represents the attribution change a faithful method should reflect, and R(τ)R(\tau) quantifies

the additional perturbation the gradient-dependent weighting introduces relative to it. Both components are informative; the model response tracks the network’s sensitivity to the perturbation, while the weight variation carries the gradient information that enables feature-level discrimination. The goal is therefore to choose τ\tau so that the gradient signal is preserved without overwhelming the model response with gradient irregularities.

The theorem therefore establishes the mechanism by which τ\tau controls weight-variation relative to the model response and the direction of this control. However, it cannot prescribe an optimal τ\tau value, because LfL_f and LgL_g are generally intractable for deep networks [26, 78] and the optimal balance between gradient information and stability is context-dependent (Figure 2). We therefore determine τ\tau empirically — see ablations in § 4.7. An exception to this are 1-Lipschitz networks [62], whose architecturally controlled constants would make R(τ)R(\tau) computable and τ\tau analytically settable, a direction we leave to future work.

By contrast, for methods such as IG where wi∝giw_i\propto g_i, gradient perturbations pass directly into the weights with an O(1)\mathcal{O}(1) Lipschitz constant and the corresponding ratio is fixed with no mechanism to control it. Moreover, integrating the weighted contributions over the full path further averages out residual gradient fluctuations, yielding even greater stability in practice than the O(1/(2τ))\mathcal{O}(1/(2\tau)) attenuation alone suggests. § 4.4 experimentally verifies this behaviour.

Experimental Evaluation

Experimental Setup

We evaluate the explanation performance of our approach across diverse architectures, namely, EfficientNet [75], MobileNetV4 [57], ResNet101 [33, 35], EvA-02 [25], DeiT3 [76], and ViT-Base [89]. Following [42, 43, 53, 72, 93], all experiments are performed on randomly sampled subsets of the ImageNet [59] and ImageNet-S [27] validation datasets using the timm [85] Python library.

We compare ERIG against four established path-based attribution techniques: Integrated Gradients (IG) [72], Important Direction Gradient Integration (IDGI) [93], Blur Integrated Gradients (BIG) [92], Guided Integrated Gradients (GIG) [43], and additionally SmoothGrad [68]. For all path-based methods, we use 200 integration steps with a black image baseline following standard practice. For ERIG, we sweep the temperature from 10−310^{-3} to 10310^{3}, with full ablations in § 4.7; while the optimum temperature can vary across architectures, in practice we find that τ=3\tau= 3 is effective across multiple tested models. For SmoothGrad, we use a 10% noise level and 50 noise samples, as prescribed.

Computationally, all the above path-based methods scale linearly with integration steps, but ERIG, IG, and IDGI have the same order of per-step runtime, whereas BIG and GIG incur substantially higher per-step overhead due to expensive gaussian blurring and sequential path construction, respectively. Comprehensive statistics in Appendix A.8.

Our evaluation employs a comprehensive suite of metrics assessing four key properties of attribution maps: (1) fidelity; (2) localization; (3) attribution stability; and (4) noise characteristics. Throughout our results, we use arrows to indicate whether higher (↑) or lower (↓) values are preferable. In all tables, the best-performing method is bold & underlined, while the second-best is bolded.

Faithfulness and Fidelity

We follow prior work [93] and adopt three metrics: Area Under the Insertion Score Curve (ISC) [53], the Accuracy Information Curve (AIC) [42, 93], and Iterative Removal of Features (DEL) [58]. ISC reveals pixels in descending attribution order and records the rise in model confidence. AIC restores pixels from a blurred baseline and tracks accuracy against image information content (MS-SSIM) [81]. DEL removes pixels in descending attribution order and tracks the decline in confidence. Higher AUC is better for ISC and AIC, lower for DEL. All three depend on the relative ranking of attribution values, so they are insensitive to attribution magnitude and to noise that does not alter the ranking. As shown in Table 1, ERIG stays comparable to the strongest baselines, with an average lead of only +0.44% across all models and metrics. The larger gains on localization, stability, and noise are consistent with this rank-only limitation, since those metrics measure properties that ranking cannot capture. Metric distributions in Appendix A.§ 3 further show substantially overlapping inter-quartile ranges between ERIG and the strongest baseline, confirming this behavior.

ModelMethodISC (↑)AIC (↑)DEL (↓)
ResNet101BIG0.31010.63480.4304
ResNet101GIG0.38120.55000.5765
ResNet101IDGI0.49230.65710.3211
ResNet101IG0.39510.68380.3884
ResNet101ERIG (ours)0.48930.66100.3270
ResNet101SmoothGrad0.18330.40110.5138
EfficientNetBIG0.37200.82410.4572
EfficientNetGIG0.38180.70610.5604
EfficientNetIDGI0.43800.84010.4310
EfficientNetIG0.35900.78820.4774
EfficientNetERIG (ours)0.45210.83230.4562
EfficientNetSmoothGrad0.25230.61220.5440
MobileNetV4BIG0.17920.66700.4574
MobileNetV4GIG0.35610.58370.5630
MobileNetV4IDGI0.28310.72230.3643
MobileNetV4IG0.16880.60230.4673
MobileNetV4ERIG (ours)0.41180.72340.3626
MobileNetV4SmoothGrad0.07650.40870.5598
ViT-BaseBIG0.56890.82730.6543
ViT-BaseGIG0.49320.68270.7949
ViT-BaseIDGI0.69170.90570.4942
ViT-BaseIG0.53540.81540.6485
ViT-BaseERIG (ours)0.69300.90130.4902
ViT-BaseSmoothGrad0.36740.57960.7975
EVA-02BIG0.53720.84200.7106
EVA-02GIG0.55220.80510.8407
EVA-02IDGI0.58760.89780.5423
EVA-02IG0.56140.87350.7255
EVA-02ERIG (ours)0.59130.89070.5373
EVA-02SmoothGrad0.38510.68760.8279
DeiT3BIG0.52800.74110.6939
DeiT3GIG0.59820.72080.7844
DeiT3IDGI0.73430.90740.5621
DeiT3IG0.64370.85140.6794
DeiT3ERIG (ours)0.73200.90100.5561
DeiT3SmoothGrad0.42920.60660.7780

Table 1. ImageNet Fidelity: ERIG places first or second on every metric and backbone, with a +0.44% average lead.

Weakly Supervised Segmentation

To evaluate the spatial correspondence between attributions and object regions, we follow previous literature [13, 36, 93] and design a weakly supervised segmentation task with ImageNet-S [27] annotations as reference. We normalize attributions to [0,1][0, 1], binarize them at thresholds t∈{0.05,0.10,…,0.95}t \in\{0.05, 0.10, \ldots, 0.95\}, and compare each binary mask against the reference mask pixel-wise. We report Intersection over Union (IoU), Pixel Accuracy (PA), and Pixel F1 Score (PF1), averaged across all thresholds. As shown in Table 2, ERIG outperforms the second-best method by ∼\sim28% on average across model-metric combinations, showcasing its effective localization capability. Unlike the fidelity metrics, the localization distributions in Appendix A.§ 4 differentiate the methods clearly, displaying a visible separation between ERIG and its baselines.

ModelMethodIoU (↑)PA (↑)PF1 (↑)
ResNet101BIG0.06430.63860.1008
ResNet101GIG0.04500.62860.0718
ResNet101IDGI0.02120.63190.0354
ResNet101IG0.07820.64010.1200
ResNet101ERIG (ours)0.16170.65800.2419
ResNet101SmoothGrad0.12860.64980.1824
EfficientNetBIG0.06290.60340.0982
EfficientNetGIG0.04480.59590.0713
EfficientNetIDGI0.01200.59370.0206
EfficientNetIG0.06390.59550.0999
EfficientNetERIG (ours)0.10850.59820.1714
EfficientNetSmoothGrad0.10120.61040.1474
MobileNetV4BIG0.06810.63570.1062
MobileNetV4GIG0.05010.62690.0797
MobileNetV4IDGI0.03050.63050.0501
MobileNetV4IG0.08340.63040.1286
MobileNetV4ERIG (ours)0.16150.65210.2424
MobileNetV4SmoothGrad0.13560.65010.1921
ViT-BaseBIG0.06560.69800.1023
ViT-BaseGIG0.03710.69460.0594
ViT-BaseIDGI0.00730.69590.0127
ViT-BaseIG0.05370.69400.0843
ViT-BaseERIG (ours)0.12030.70540.1893
ViT-BaseSmoothGrad0.11500.71130.1635
EVA-02BIG0.05400.69580.0856
EVA-02GIG0.03960.69010.0635
EVA-02IDGI0.00740.69470.0128
EVA-02IG0.05290.68460.0838
EVA-02ERIG (ours)0.13170.70540.2031
EVA-02SmoothGrad0.10380.69520.1526
DeiT3BIG0.04700.65110.0746
DeiT3GIG0.01730.64600.0279
DeiT3IDGI0.00670.64760.0116
DeiT3IG0.02060.64670.0329
DeiT3ERIG (ours)0.11790.66370.1819
DeiT3SmoothGrad0.02960.64510.0438

Table 2. ImageNet-S Localization: ERIG improves localization by ∼28% on average over the next-best method.

Attribution Stability

In deployment, attribution maps should remain stable under small input perturbations. We evaluate this with three established metrics, Max Sensitivity (MS) [94], Local Lipschitz Constant (LLC) [5], and Relative Input Stability (RIS) [4]. These quantify the worst-case attribution shift within an ℓ∞\ell_{\infty} ball of radius rr; lower is more stable.

MS⁡(x,r)=max⁡∥x−x′∥∞≤r∥ex−ex′∥2(8)\operatorname{MS}(\mathbf{x},r)=\max_{\lVert\mathbf{x}-\mathbf{x}'\rVert_{\infty}\le r}\lVert\mathbf{e}_{\mathbf{x}}-\mathbf{e}_{\mathbf{x}'}\rVert_{2} \tag*{(8)}
LLC⁡(x,r)=max⁡∥x−x′∥∞≤r∥ex−ex′∥2∥x−x′∥2(9)\operatorname{LLC}(\mathbf{x},r)=\max_{\lVert\mathbf{x}-\mathbf{x}'\rVert_{\infty}\le r}\frac{\lVert\mathbf{e}_{\mathbf{x}}-\mathbf{e}_{\mathbf{x}'}\rVert_{2}}{\lVert\mathbf{x}-\mathbf{x}'\rVert_{2}} \tag*{(9)}
RIS⁡(x,r)=max⁡∥x−x′∥∞≤r∥ex−ex′∥2/∥ex∥2∥x−x′∥2/∥x∥2(10)\operatorname{RIS}(\mathbf{x},r)=\max_{\lVert\mathbf{x}-\mathbf{x}'\rVert_{\infty}\le r}\frac{\lVert\mathbf{e}_{\mathbf{x}}-\mathbf{e}_{\mathbf{x}'}\rVert_{2}/\lVert\mathbf{e}_{\mathbf{x}}\rVert_{2}}{\lVert\mathbf{x}-\mathbf{x}'\rVert_{2}/\lVert\mathbf{x}\rVert_{2}} \tag*{(10)}

Following their protocols [4, 5, 94], we approximate each supremum via Monte Carlo sampling of 50 isotropic uniform perturbations per image at r=0.1r = 0.1 over 100 ImageNet validation samples; isotropic perturbations introduce no directional bias and assess local smoothness uniformly across all input directions. Appendix A.15 complements this protocol with adversarially selected perturbation directions, under which ERIG’s stability advantage carries over unchanged. As reported in Table 3, ERIG outperforms the best baseline method by up to two orders of magnitude, confirming its superior attribution map stability. Log-scale plots with inter-quartile ranges in Appendix A.§ 5 further confirm their significance.

ModelMethodMS (↓)LLC (↓)RIS (↓)
ResNet101BIG0.17500.007812.498
ResNet101GIG0.37230.016688.373
ResNet101IDGI0.01000.000418.653
ResNet101IG0.68720.030730.295
ResNet101ERIG (ours)8.6e−053.8e−061.066
ResNet101SmoothGrad0.13910.0062102.340
EfficientNetBIG0.03160.00106.318
EfficientNetGIG0.23000.007223.019
EfficientNetIDGI0.00400.000112.666
EfficientNetIG0.34630.010818.834
EfficientNetERIG (ours)6.7e−052.1e−061.271
EfficientNetSmoothGrad0.04580.001416.972
MobileNetV4BIG0.16380.00649.863
MobileNetV4GIG0.55140.021530.213
MobileNetV4IDGI0.00510.000215.979
MobileNetV4IG0.90210.035223.123
MobileNetV4ERIG (ours)1.0e−045.4e−062.196
MobileNetV4SmoothGrad0.11660.004424.148
ViT-BaseBIG0.06520.00177.142
ViT-BaseGIG0.21670.005620.346
ViT-BaseIDGI0.00930.000215.883
ViT-BaseIG0.46780.012220.596
ViT-BaseERIG (ours)3.0e−049.0e−064.946
ViT-BaseSmoothGrad0.04520.001212.456
EVA-02BIG0.04830.001111.652
EVA-02GIG0.15860.003521.562
EVA-02IDGI0.00530.000113.423
EVA-02IG0.65620.0146134.160
EVA-02ERIG (ours)6.7e−051.5e−061.601
EVA-02SmoothGrad0.04380.001016.770
DeiT3BIG0.09670.00259.570
DeiT3GIG0.17460.004549.933
DeiT3IDGI0.00780.000213.132
DeiT3IG0.64090.016742.144
DeiT3ERIG (ours)3.9e−051.0e−060.756
DeiT3SmoothGrad0.03800.001040.339

Table 3. ImageNet Stability: ERIG improves stability by ∼55× on average over the next-best method.

Quantitative Estimation of Attribution Noise

A critical limitation of stability metrics such as MS, LLC, and RIS is their high computational cost, as they require multiple perturbed inputs and the corresponding attribution maps. For path-based methods with TT integration steps, evaluating 100 images with 50 perturbations each requires 50×100×T50 \times100 \times T forward and backward passes, reaching a million evaluations for T=200T = 200. Since attribution noise and instability both arise from irregularities in the gradient field and sampling noise from the input image, attribution noise level serves as an efficient proxy for stability. We hence adapt two established image-noise estimators that directly quantify attribution noise through principled statistical estimation. These metrics also address a limitation of prior work, which largely assessed noise reduction qualitatively through visual inspection or anecdotal examples rather than direct quantitative measurement [43, 93]. These estimators enable objective and reproducible comparisons across attribution methods. However, we would like to categorically mention that they should be used along with fidelity and localization metrics, as reduced noise and improved stability alone do not guarantee explanation faithfulness or accurate localization. In ERIG, we balance these multiple competing factors using an optimum temperature (τ\tau) value. For a comparison between LINV and other complexity, compressibility, and sparsity measures, see Appendix A.14.

Convolution-Based Noise Estimation (LINV-C) [39]

σ2=136(H−2)(W−2)∑i=2H−1∑j=2W−1(Aij∗S)2(11)\sigma^2 = \frac{1}{36(H-2)(W-2)} \sum_{i=2}^{H-1} \sum_{j=2}^{W-1} \left(A_{ij} \ast S\right)^2 \tag*{(11)}

LINV-C detects unstructured variations characteristic of random perturbations. In Eq. (11), A∈RH×WA \in\mathbb{R}^{H \times W} is the attribution map, ∗\ast denotes convolution, and SS is the 3×33 \times3 noise-isolating filter with rows (1,−2,1)(1, -2, 1), (−2,4,−2)(-2, 4, -2), (1,−2,1)(1, -2, 1); the resulting variance σ2\sigma^2 estimates how unpredictably attribution values fluctuate across the map.

PCA-Based Noise Estimation (LINV-P) [16] LINV-P exploits the separation between low-rank structure and Gaussian noise in local attribution patches. Each map is partitioned into non-overlapping patches xt∈Rrx_t \in\mathbb{R}^{r}, modeled as xt=Ayt+etx_t = Ay_t + e_t with et∼Nr(0,σ2I)e_t \sim\mathcal{N}_r(0, \sigma^2 I), where A∈Rr×mA \in\mathbb{R}^{r \times m} spans the principal subspace capturing coherent structure. The eigenvalues λi\lambda_i of the empirical patch covariance Xs=(xt)t=1sX_s = (x_t)_{t=1}^{s} separate into structure and noise components, with the noise eigenvalues asymptotically following λi∼N(σ2,2σ4/s)\lambda_i \sim\mathcal{N}(\sigma^2, 2\sigma^4/s) for i∈[m+1,r]i \in[m+1,r]. We iteratively

remove dominant eigenvalues until the mean of the remaining ones equals their median, yielding a robust estimate of the noise variance.

Both estimators convert attribution noise into a quantitative, reproducible estimate. For interpretability, we report the Log-Inverse Noise Variance (LINV-X), defined as −log⁡10(σ2)-\log_{10}(\sigma^2). Here higher values indicate cleaner, more stable attribution maps.

Under these metrics, ERIG achieves attribution noise levels several orders lower than competing methods (Table 4), and the MobileNetV4 distributions in Figure 4 show a clear ordering. GIG [43], which steers the path away from high-gradient regions, outperforms vanilla IG but trails SmoothGrad’s [68] explicit perturbation averaging. IDGI [93] in turn surpasses SmoothGrad by weighting contributions with squared gradients. ERIG outperforms all three, so its entropy-regularized weighting attenuates noise more effectively than path-steering, external smoothing, and direct gradient weighting. The same ordering appears in the stability results (Table 3), supporting LINV as a meaningful proxy for stability. LINV thus adds a noise dimension absent from previous evaluation protocols; together with the fidelity, localization, and stability metrics, it enables a more complete evaluation of attribution quality.

ModelMethodLINV-C (↑\uparrow)LINV-P (↑\uparrow)
ResNet101BIG6.64906.6053
ResNet101GIG7.96547.9093
ResNet101IDGI9.43089.4017
ResNet101IG6.58696.4886
ResNet101ERIG (ours)16.806916.7781
ResNet101SmoothGrad10.13559.9179
EfficientNetBIG7.91658.0544
EfficientNetGIG7.36407.2324
EfficientNetIDGI10.242910.2457
EfficientNetIG6.87666.8567
EfficientNetERIG (ours)20.704820.7121
EfficientNetSmoothGrad8.72878.7101
MobileNetV4BIG6.50566.4759
MobileNetV4GIG6.59816.4472
MobileNetV4IDGI10.086310.0732
MobileNetV4IG5.97125.7927
MobileNetV4ERIG (ours)18.130318.1133
MobileNetV4SmoothGrad7.84767.6511
ViT-BaseBIG8.03627.9448
ViT-BaseGIG8.00557.9185
ViT-BaseIDGI10.161910.0085
ViT-BaseIG7.22257.0610
ViT-BaseERIG (ours)19.342619.2005
ViT-BaseSmoothGrad9.25998.8763
EVA-02BIG8.46748.4218
EVA-02GIG8.50478.3316
EVA-02IDGI10.802710.7783
EVA-02IG7.56777.5228
EVA-02ERIG (ours)20.128520.1150
EVA-02SmoothGrad9.59539.1273
DeiT3BIG7.74678.1019
DeiT3GIG8.65768.7212
DeiT3IDGI10.247110.2542
DeiT3IG7.15757.5698
DeiT3ERIG (ours)19.577819.5845
DeiT3SmoothGrad9.59209.3614

Table 4. ImageNet Noise: ERIG attains the highest LINV on every backbone, 6.7–10.5 LINV above the next best.

Figure 4. MobileNetV4 Noise Histograms

Generalization Beyond ImageNet

To verify that our findings are not specific to the ImageNet family, we evaluate the full protocol on Oxford-IIIT Pet [55], a fine-grained dataset with segmentation trimaps, using fine-tuned ResNet-50, MobileNetV2, ConvNeXt-B, ViT-B, and DINOv2-S checkpoints. We also evaluate on the FunnyBirds benchmark [37] with ResNet-50, VGG16, and ViT-B backbones. The ImageNet conclusions carry over to both settings. On Oxford Pet, ERIG remains competitive with the strongest baseline on ranking-based fidelity (mean gap of −0.40%-0.40\%), leads localization by 20.2% on average, improves stability by ∼\sim38×\times on average over the next-best method, and achieves attribution noise levels several orders lower than competing methods. On FunnyBirds, ERIG leads completeness by 3.4% and contrastivity by 2.1% on average while staying within 1.5% of the strongest baseline on correctness. Full analysis in Appendix A.11.

Ablation Studies

Effect of Temperature on Fidelity and Localization. As shown in Figure 5a and Figure 5b, fidelity and localization peak at an intermediate τ\tau. At low temperatures, the distribution is over-concentrated on a few dominant and possibly noisy gradient components, causing attributions to miss broader informative structure. At high temperatures, the distribution approaches uniformity and loses discriminability between informative and uninformative features. The optimal temperature (τ≈3\tau\approx3) balances attribution sharpness against over-concentration, consistent with the balance characterized by R(τ)R(\tau) in Theorem 2.

Figure 5. Ablation of softmax temperature τ\tau in ERIG across ResNet101, ViT-Base, and EVA02 models. All plots use a log-scale x-axis and a linear-scale y-axis, except stability, which also uses a log-scale y-axis.

Effect of Temperature on Stability and Noise Reduction. Figure 5c shows that for τ<3\tau< 3, attributions exhibit substantially higher distortion under input perturbations, as the concentrated softmax amplifies spurious gradient spikes that destabilize attribution maps. Beyond this region, distortion drops sharply and plateaus as the entropy objective redistributes attribution mass over a broader, more stable feature set. Concurrently, Figure 5d shows the inverse noise metric increasing monotonically with τ\tau, where a more uniform softmax prevents any single high-gradient component from dominating and dilutes the influence of noise spikes. Together, these observations mark τ≈3\tau\approx3 as a sweet spot, where stability and noise suppression are sufficient while fidelity and localization remain near their peak.

Complementarity with SmoothGrad. Additionally, we validate our findings by ensembling all methods with SmoothGrad in Appendix A.10. There, stability and noise metrics follow the same patterns, confirming that ERIG’s intrinsic noise attenuation properties are complementary to external smoothing provided by ensemble methods.

Baseline Sensitivity. Finally, we also ablate the baseline choice, evaluating ERIG across twelve alternatives on EVA02 and ResNet101 in Appendix A.13. Constant references track the zero default closely, other families less so; since all methods share one baseline, the relative claims hold.

Conclusion

We present Entropy-Regularized Integrated Gradients (ERIG), a path-based attribution method that converts squared-gradient sensitivities into attribution weights using an entropy-regularized softmax. A temperature parameter τ\tau balances sensitivity preservation against attenuation of local gradient irregularities. Our Lipschitz analysis decomposes attribution perturbations into a model-response term with a τ\tau-independent bound and a weight-variation term whose bound decays as O(1/(2τ))O(1/(2\tau)). Their ratio, R(τ)R(\tau), characterizes the trade-off between gradient information and attribution stability. We also adapt established noise estimators into LINV-C and LINV-P. Across eleven models on ImageNet and Oxford-IIIT Pet, ERIG achieves fidelity comparable to the strongest baselines while substantially improving localization, stability, and noise suppression, all at a per-step cost within ∼\sim5% of IDGI and ∼\sim25% of IG. On Funny-Birds, it places first or second in all metrics. Future work will explore further adaptive temperature schedules τ(t)\tau(t), analytic τ\tau selection on 1-Lipschitz networks, and composition with gradient-flow corrections such as LibraGrad.

Acknowledgements

The authors acknowledge NSF (IIS-2331908 and OAC-2530965), DOE (DE-SC0024576), DARPA (HR00112420004 and HR00112490420), NAIRR Pilot (NAIRR250261), Indiana Jetstream2 GPU (CIS240951 via ACCESS [10, 31]), OpenAI, Google Cloud, and NCSA Delta GPU for partially contributing to this research. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of funding agencies and companies mentioned above.

Proof of Local Sensitivity

Let ε\varepsilon be a very small (ε≪1\varepsilon\ll1) random unbiased (E[ε]=0\mathbf{E}[\varepsilon] = 0) perturbation of the ii-th coordinate of input x\boldsymbol{x}. The corresponding perturbed vector is xε={x1,x2,…,xi+ε,…,xN}\boldsymbol{x}_{\varepsilon} = \{x_1, x_2, \ldots, x_i + \varepsilon, \ldots, x_N\}. A Taylor expansion of ff around x\boldsymbol{x} in the ii-th coordinate gives

f(xε)=f(x)+∂f∂xi⋅ε+12∂2f∂xi2⋅ε2+⋯=f(x)+∂f∂xi⋅ε+O(ε2)(12)\begin{aligned} f(\boldsymbol{x}_{\varepsilon}) &= f(\boldsymbol{x}) + \frac{\partial f}{\partial x_i} \cdot\varepsilon+ \frac{1}{2} \frac{\partial^2 f}{\partial x_i^2} \cdot\varepsilon^2 + \cdots\\ &= f(\boldsymbol{x}) + \frac{\partial f}{\partial x_i} \cdot\varepsilon+ \mathcal{O}(\varepsilon^2) \tag*{(12)} \end{aligned}

Taking expectation of f(xε)f(\boldsymbol{x}_{\varepsilon}) gives us,

E[f(xε)]=E[f(x)+∂f∂xi⋅ε+O(ε2)](13)\mathbf{E}[f(\boldsymbol{x}_{\varepsilon})] = \mathbf{E}\left[f(\boldsymbol{x}) + \frac{\partial f}{\partial x_i} \cdot\varepsilon+ \mathcal{O}(\varepsilon^2)\right] \tag*{(13)}
=f(x)+∂f∂xi⋅E[ε]+O(ε2)(14)= f(\boldsymbol{x}) + \frac{\partial f}{\partial x_i} \cdot\mathbf{E}[\varepsilon] + \mathcal{O}(\varepsilon^2) \tag*{(14)}
=f(x)+O(ε2)(15)= f(\boldsymbol{x}) + \mathcal{O}(\varepsilon^2) \tag*{(15)}

Hence, we can compute the variance as,

Var⁡[f(xε)]=E[(f(xε)−E[f(xε)])2]=E[(∂f∂xi⋅ε+O(ε2))2](16)\begin{aligned} \operatorname{Var}[f(\boldsymbol{x}_{\varepsilon})] &= \mathbf{E}\left[\left(f(\boldsymbol{x}_{\varepsilon}) - \mathbf{E}[f(\boldsymbol{x}_{\varepsilon})]\right)^2\right] \\ &= \mathbf{E}\left[\left(\frac{\partial f}{\partial x_i} \cdot\varepsilon+ \mathcal{O}(\varepsilon^2)\right)^2\right] \tag*{(16)} \end{aligned}
=(∂f∂xi)2⋅E[ε2]+O(ε3)(17)= \left(\frac{\partial f}{\partial x_i}\right)^2 \cdot\mathbf{E}[\varepsilon^2] + \mathcal{O}(\varepsilon^3) \tag*{(17)}
∝(∂f∂xi)2(18)\propto\left(\frac{\partial f}{\partial x_i}\right)^2 \tag*{(18)}

Here, E[ε2]\mathbf{E}[\varepsilon^2] acts as the proportionality constant in the approximation, while the squared partial derivative, gi2=(∂f/∂xi)2g_i^2 = (\partial f/\partial x_i)^2, measures the local sensitivity of the function ff to feature xix_i. This quantity, integrated over the entire input space, forms the basis of Derivative-based Global Sensitivity Measures (DGSMs) [46, 70], a widely used alternative to Sobol’ sensitivity indices [69]. DGSMs have been applied across a range of domains, including input screening in environmental models [6], sensitivity analysis of ODEs in biology [98], and uncertainty quantification in geophysical simulations [11], etc.

Maximum Entropy Weight Derivation

Solution via KKT Conditions

We address the constrained optimization problem via the Karush-Kuhn-Tucker (KKT) conditions. First, we restate the problem in the following standard form:

min⁡w∈[0,1]N(∑i=1Nwiln⁡wi),s.t.(19)\min_{\boldsymbol{w} \in[0,1]^N} \left(\sum_{i=1}^{N} w_i \ln w_i\right), \qquad\text{s.t.} \tag*{(19)}
(1−∑i=1Nwi)=0,(γ−∑i=1Nwigi2)≤0(20)\left(1 - \sum_{i=1}^{N} w_i\right) = 0, \quad\left(\gamma- \sum_{i=1}^{N} w_i g_i^2\right) \leq0 \tag*{(20)}

Next, we construct the Lagrangian:

L(w,α,λ)=∑i=1Nwiln⁡wi+α(1−∑i=1Nwi)+λ(γ−∑i=1Nwigi2)(21)\mathcal{L}(\boldsymbol{w}, \alpha, \lambda) = \sum_{i=1}^{N} w_i \ln w_i + \alpha\left(1 - \sum_{i=1}^{N} w_i\right) + \lambda\left(\gamma- \sum_{i=1}^{N} w_i g_i^2\right) \tag*{(21)}

where α∈R\alpha\in\mathbb{R} is the Lagrange multiplier for the equality constraint (normalization), and λ≥0\lambda\geq0 is the Lagrange multiplier for the inequality constraint (gradient magnitude threshold). The KKT conditions for optimality are:

Stationarity:∂L∂wi=1+ln⁡wi−α−λgi2=0,∀i∈[N],(22)\text{Stationarity:}\qquad\frac{\partial\mathcal{L}}{\partial w_i} = 1 + \ln w_i - \alpha- \lambda g_i^2 = 0, \forall i \in[N], \tag*{(22)}

Primal feasibility:

∑i=1Nwi=1,∑i=1Nwigi2≥γ,(23)\sum_{i=1}^{N} w_i = 1,\qquad\sum_{i=1}^{N} w_i g_i^2 \ge\gamma, \tag*{(23)}

Dual feasibility:

λ≥0,(24)\lambda\ge0, \tag*{(24)}

Complementary slackness:

λ(γ−∑i=1Nwigi2)=0.(25)\lambda\left(\gamma-\sum_{i=1}^{N}w_i g_i^2\right)=0. \tag*{(25)}

From the stationarity condition (22), we can solve for wiw_i:

ln⁡wi=(α−1)+λgi2,(26)\ln w_i = (\alpha-1)+\lambda g_i^2, \tag*{(26)}
wi=exp⁡(α−1)exp⁡(λgi2).(27)w_i=\exp(\alpha-1)\exp(\lambda g_i^2). \tag*{(27)}

Using the normalization constraint ∑i=1Nwi=1\sum_{i=1}^{N}w_i=1 from the primal feasibility conditions (23), we eliminate the constant exp⁡(α−1)\exp(\alpha-1):

∑i=1Nwi=exp⁡(α−1)∑i=1Nexp⁡(λgi2)=1,(28)\sum_{i=1}^{N}w_i=\exp(\alpha-1)\sum_{i=1}^{N}\exp(\lambda g_i^2)=1, \tag*{(28)}
exp⁡(α−1)=1∑j=1Nexp⁡(λgj2).(29)\exp(\alpha-1)=\frac{1}{\sum_{j=1}^{N}\exp(\lambda g_j^2)}. \tag*{(29)}

Substituting equation (29) into equation (27), we obtain the normalized weights:

wi∗=exp⁡(λgi2)∑j=1Nexp⁡(λgj2)=Softmax⁡i(λg2),(30)w_i^{*}=\frac{\exp(\lambda g_i^2)}{\sum_{j=1}^{N}\exp(\lambda g_j^2)}=\operatorname{Softmax}_i(\lambda\mathbf{g}^2), \tag*{(30)}

where Softmax⁡i(⋅)\operatorname{Softmax}_i(\cdot) denotes the ii-th component of the softmax function.

The Lagrange multiplier λ≥0\lambda\ge0 controls the trade-off between entropy maximization and gradient emphasis. When λ=0\lambda=0, we recover the uniform distribution (maximum entropy with no gradient bias). As λ\lambda increases, the distribution concentrates more on features with larger gradient magnitudes while assigning lower probability to features with smaller gradients.

Relationship Between λ\lambda and γ\gamma

From the complementary slackness condition (25), at the optimal solution either λ=0\lambda=0 (in which case the gradient strength constraint will not be active resulting in a uniform distribution) or the constraint holds with equality:

∑i=1Nwi∗gi2=γ.(31)\sum_{i=1}^{N}w_i^{*}g_i^2=\gamma. \tag*{(31)}

Substituting Eq. (30) into Eq. (31), we obtain the relationship between λ\lambda and γ\gamma:

∑i=1Ngi2exp⁡(λgi2)∑j=1Nexp⁡(λgj2)=γ.(32)\sum_{i=1}^{N}\frac{g_i^2\exp(\lambda g_i^2)}{\sum_{j=1}^{N}\exp(\lambda g_j^2)}=\gamma. \tag*{(32)}

This equation implicitly defines λ\lambda as a function of γ\gamma. In principle, for a given threshold γ\gamma, one could solve Eq. (32) numerically to find the corresponding λ\lambda.

Practical Implementation

While tuning γ\gamma would require solving the constrained optimization problem for each value, it is more practical to tune λ\lambda directly as a hyperparameter. We denote this hyperparameter as τ=1/λ\tau=1/\lambda (the temperature parameter in the softmax), which provides direct control over the concentration of the distribution on high-gradient regions without requiring iterative optimization. Hence, the ERIG weighting scheme becomes:

wi∗(g(x);τ)=Softmax⁡i(∇f⊙∇fτ)=exp⁡(gi2/τ)∑j=1Nexp⁡(gj2/τ).(33)w_i^{*}(\mathbf{g}(\mathbf{x});\tau)=\operatorname{Softmax}_i\left(\frac{\nabla f\odot\nabla f}{\tau}\right)=\frac{\exp(g_i^2/\tau)}{\sum_{j=1}^{N}\exp(g_j^2/\tau)}. \tag*{(33)}

More Visualization Examples

The attribution values are mapped to a [0,1][0,1] range following the methodology of Integrated Gradients [72]. Below, you can find more comparisons of attribution maps for the MobileNetV4 model:

Figure 6. Additional visualizations of attribution maps for the MobileNetV4 architecture.

Fidelity Distributions

To complement the fidelity results in Table 1, we plot the per-sample distributions of each fidelity metric across all six architectures. Each shaded region spans the inter-quartile range (25th to 75th percentile); the horizontal line inside the shaded region represents the mean, and the vertical capped lines indicate the full range across samples. As discussed in § 4.2, the substantial overlap in inter-quartile ranges across methods is consistent with the rank-only nature of these metrics, since they depend solely on attribution ranking and methods with markedly different attribution magnitudes and noise characteristics can produce similar fidelity scores.

Figure 7. ISC distribution with inter-quartile range and full range across all architectures. Higher is better.

Figure 8. AIC distribution with inter-quartile range and full range across all architectures. Higher is better.

DEL distribution across all architectures

Figure 9. DEL distribution with inter-quartile range and full range across all architectures. Lower is better.

Localization Distributions

Here we plot the per-sample distributions of each localization metric in Table 2 across all six architectures using the same format as Sec. A.4. In contrast to the fidelity distributions, ERIG’s inter-quartile ranges are visibly higher than those of all baselines for IoU and Pixel F1, confirming the localization advantages reported in the main text. Pixel Accuracy shows less separation because, unlike IoU and Pixel F1, it does not penalize false positives relative to the object region and is therefore less sensitive to differences in localization quality. This limitation is well known and has been studied by Csurka et al. [18] in 2013 and, more recently, by Wang et al. [82] in 2023.

IoU distribution across all architectures

Figure 10. IoU distribution with inter-quartile range and full range across all architectures. Higher is better.

Figure 11. Pixel Accuracy distribution with inter-quartile range and full range across all architectures. Higher is better.

Figure 12. Pixel F1 distribution with inter-quartile range and full range across all architectures. Higher is better.

Stability Distributions

To complement the tabulated stability results in Table 3, we plot each stability metric across all six architectures on a log⁡10\log_{10} scale for ease of visualization. Each shaded region spans the inter-quartile range (25th to 75th percentile); the horizontal line inside the shaded region represents the mean, and the vertical capped lines indicate the full range across samples. The ERIG bar is visually highlighted for clarity. Across all metrics and models, ERIG’s inter-quartile ranges are entirely non-overlapping with those of every baseline, confirming that the improvements reported in the main text are meaningful.

Figure 13. Max Sensitivity (MS) with inter-quartile range and full range across all architectures (log⁡10\log_{10} scale). Lower is better.

Figure 14. Local Lipschitz Constant (LLC) with inter-quartile range and full range across all architectures (log⁡10\log_{10} scale). Lower is better.

Figure 15. Relative Input Stability (RIS) with inter-quartile range and full range across all architectures (log⁡10\log_{10} scale). Lower is better.

Proof of Lipschitz Decomposition

Let Γ(t)=xb+td\Gamma(t) = \mathbf{x}^{b} + t\mathbf{d} with d=xa−xb\mathbf{d} = \mathbf{x}^{a} - \mathbf{x}^{b} and Γ~(t)=Γ(t)+tη\widetilde{\Gamma}(t) = \Gamma(t) + t\boldsymbol{\eta}. We bound ∥M∥2\|\mathbf{M}\|_{2} and ∥W∥2\|\mathbf{W}\|_{2} separately.

Model-Response Term

Since wi∗∈[0,1]w_i^* \in[0,1] and ∑iwi∗=1\sum_i w_i^* = 1,

∥M∥2=∥∫01w∗(g(t);τ) Δf˙(t) dt∥2≤∫01∣Δf˙(t)∣ dt.(34)\|\mathbf{M}\|_2 = \left\|\int_0^1 \mathbf{w}^*(\mathbf{g}(t);\tau)\,\Delta\dot{f}(t)\,dt\right\|_2 \le\int_0^1 |\Delta\dot{f}(t)|\,dt. \tag*{(34)}

Define ϕ(t)=f(Γ~(t))−f(Γ(t))\phi(t) = f(\widetilde{\Gamma}(t)) - f(\Gamma(t)), so that Δf˙(t)=ϕ′(t)\Delta\dot{f}(t) = \phi'(t). The derivative of ϕ\phi is the difference of the directional derivatives of ff along Γ~(t)\widetilde{\Gamma}(t) and Γ(t)\Gamma(t):

ϕ′(t)=∇f(Γ(t)+tη)⋅(d+η)−∇f(Γ(t))⋅d(35)\phi'(t) = \nabla f(\Gamma(t) + t\boldsymbol{\eta}) \cdot(\mathbf{d} + \boldsymbol{\eta}) - \nabla f(\Gamma(t)) \cdot\mathbf{d} \tag*{(35)}
=[∇f(Γ(t)+tη)−∇f(Γ(t))]⋅d+∇f(Γ(t)+tη)⋅η.(36)= \left[\nabla f(\Gamma(t) + t\boldsymbol{\eta}) - \nabla f(\Gamma(t))\right] \cdot\mathbf{d} + \nabla f(\Gamma(t) + t\boldsymbol{\eta}) \cdot\boldsymbol{\eta}. \tag*{(36)}

Using ∥Γ(t)+tη−Γ(t)∥2=t∥η∥2\|\Gamma(t) + t\boldsymbol{\eta} - \Gamma(t)\|_2 = t\|\boldsymbol{\eta}\|_2, ∥d∥2=∣Γ∣\|\mathbf{d}\|_2 = |\Gamma| with Assumptions 1 and 2, one can write,

∣ϕ′(t)∣≤Lgt∥η∥2∣Γ∣+Lf∥η∥2.(37)|\phi'(t)| \le L_g t\|\boldsymbol{\eta}\|_2|\Gamma| + L_f\|\boldsymbol{\eta}\|_2. \tag*{(37)}

Integrating over [0,1][0,1] yields the model-response bound,

∥M∥2≤∫01(Lgt∥η∥2∣Γ∣+Lf∥η∥2) dt=(Lf+12Lg∣Γ∣)∥η∥2.(38)\|\mathbf{M}\|_2 \le\int_0^1 \left(L_g t\|\boldsymbol{\eta}\|_2|\Gamma| + L_f\|\boldsymbol{\eta}\|_2\right)\,dt = \left(L_f + \frac{1}{2}L_g|\Gamma|\right)\|\boldsymbol{\eta}\|_2. \tag*{(38)}

Weight-Variation Term

We first bound the sensitivity of the softmax weights to gradient perturbations. Let g∈RN\mathbf{g} \in\mathbb{R}^{N} and g~=g+ϵ\widetilde{\mathbf{g}} = \mathbf{g} + \boldsymbol{\epsilon} denote the gradient vectors at Γ(t)\Gamma(t) and Γ~(t)\widetilde{\Gamma}(t) respectively, and let ⊙\odot denote the Hadamard product. From [51], the softmax operator is (1/(2τ))(1/(2\tau))-Lipschitz with respect to any ℓp\ell_p norm. Therefore,

∥Softmax⁡(g~2τ)−Softmax⁡(g2τ)∥2≤12τ∥g~2−g2∥2.(39)\left\lVert\operatorname{Softmax}\left(\frac{\widetilde{\boldsymbol{g}}^2}{\tau}\right)-\operatorname{Softmax}\left(\frac{\boldsymbol{g}^2}{\tau}\right)\right\rVert_2 \le\frac{1}{2\tau}\left\lVert\widetilde{\boldsymbol{g}}^2-\boldsymbol{g}^2\right\rVert_2. \tag*{(39)}

Using g~=g+ϵ\widetilde{\boldsymbol{g}}=\boldsymbol{g}+\boldsymbol{\epsilon}, we expand elementwise,

g~2−g2=(g+ϵ)2−g2=2ϵ⊙g+ϵ2.(40)\begin{aligned} \widetilde{\boldsymbol{g}}^2-\boldsymbol{g}^2 &= (\boldsymbol{g}+\boldsymbol{\epsilon})^2-\boldsymbol{g}^2 \\ &=2\boldsymbol{\epsilon}\odot\boldsymbol{g}+\boldsymbol{\epsilon}^2. \tag*{(40)} \end{aligned}

Next,

∥g~2−g2∥2=∥2ϵ⊙g+ϵ2∥2≤2∥ϵ⊙g∥2+∥ϵ2∥2.(41)\begin{aligned} \left\lVert\widetilde{\boldsymbol{g}}^2-\boldsymbol{g}^2\right\rVert_2 &= \left\lVert2\boldsymbol{\epsilon}\odot\boldsymbol{g}+\boldsymbol{\epsilon}^2\right\rVert_2 \\ &\le2\left\lVert\boldsymbol{\epsilon}\odot\boldsymbol{g}\right\rVert_2+\left\lVert\boldsymbol{\epsilon}^2\right\rVert_2. \tag*{(41)} \end{aligned}

Since each gi2≤∥g∥22g_i^2\le\lVert\boldsymbol{g}\rVert_2^2, we have

∥ϵ⊙g∥2=∑i=1Nϵi2gi2≤∥g∥22∑i=1Nϵi2=∥ϵ∥2∥g∥2.(42)\begin{aligned} \left\lVert\boldsymbol{\epsilon}\odot\boldsymbol{g}\right\rVert_2=\sqrt{\sum_{i=1}^{N}\epsilon_i^2g_i^2} \\ &\le\sqrt{\lVert\boldsymbol{g}\rVert_2^2\sum_{i=1}^{N}\epsilon_i^2}=\lVert\boldsymbol{\epsilon}\rVert_2\lVert\boldsymbol{g}\rVert_2. \tag*{(42)} \end{aligned}

Similarly, ∥ϵ2∥2=∑iϵi4≤∑iϵi2=∥ϵ∥22\lVert\boldsymbol{\epsilon}^2\rVert_2=\sqrt{\sum_i\epsilon_i^4}\le\sum_i\epsilon_i^2=\lVert\boldsymbol{\epsilon}\rVert_2^2. Combining,

∥Softmax⁡(g~2τ)−Softmax⁡(g2τ)∥2≤2∥ϵ∥2∥g∥2+∥ϵ∥222τ.(43)\left\lVert\operatorname{Softmax}\left(\frac{\widetilde{\boldsymbol{g}}^2}{\tau}\right)-\operatorname{Softmax}\left(\frac{\boldsymbol{g}^2}{\tau}\right)\right\rVert_2\le\frac{2\lVert\boldsymbol{\epsilon}\rVert_2\lVert\boldsymbol{g}\rVert_2+\lVert\boldsymbol{\epsilon}\rVert_2^2}{2\tau}. \tag*{(43)}

We apply this to path parameter tt by defining ϵ(t)=∇f(Γ(t)+tη)−∇f(Γ(t))\boldsymbol{\epsilon}(t)=\nabla f(\Gamma(t)+t\boldsymbol{\eta})-\nabla f(\Gamma(t)). Then, Assumption 1 gives ∥g(t)∥2≤Lf\lVert\boldsymbol{g}(t)\rVert_2\le L_f and Assumption 2 gives ∥ϵ(t)∥2≤Lgt∥η∥2\lVert\boldsymbol{\epsilon}(t)\rVert_2\le L_g t\lVert\boldsymbol{\eta}\rVert_2. So, the pointwise weight perturbation satisfies

∥Δw∗(t)∥2≤LfLgt∥η∥2τ+O(∥η∥22).(44)\lVert\Delta\boldsymbol{w}^{*}(t)\rVert_2\le\frac{L_fL_gt\lVert\boldsymbol{\eta}\rVert_2}{\tau}+\mathcal{O}(\lVert\boldsymbol{\eta}\rVert_2^2). \tag*{(44)}

The function increment rate along the unperturbed path satisfies ∣f˙(t)∣=∣∇f(Γ(t))⋅d∣≤Lf∣Γ∣|\dot{f}(t)|=|\nabla f(\Gamma(t))\cdot\boldsymbol{d}|\le L_f|\Gamma| by Assumption 1. Integrating the product over [0,1][0,1] yields the weight-variation bound,

∥W∥2≤∫01∥Δw∗(t)∥2∣f˙(t)∣ dt≤Lf2Lg∣Γ∣∥η∥2τ∫01t dt+O(∥η∥22)=Lf2Lg∣Γ∣∥η∥22τ+O(∥η∥22).(45)\lVert\boldsymbol{W}\rVert_2\le\int_0^1\lVert\Delta\boldsymbol{w}^{*}(t)\rVert_2|\dot{f}(t)|\,dt\le\frac{L_f^2L_g|\Gamma|\lVert\boldsymbol{\eta}\rVert_2}{\tau}\int_0^1t\,dt+\mathcal{O}(\lVert\boldsymbol{\eta}\rVert_2^2)=\frac{L_f^2L_g|\Gamma|\lVert\boldsymbol{\eta}\rVert_2}{2\tau}+\mathcal{O}(\lVert\boldsymbol{\eta}\rVert_2^2). \tag*{(45)}

Applying the triangle inequality to ΔERIG=M+W\Delta\mathrm{ERIG}=\boldsymbol{M}+\boldsymbol{W},

∥ΔERIG∥2=∥M+W∥2≤∥M∥2+∥W∥2(46)\lVert\Delta\mathrm{ERIG}\rVert_2=\lVert\boldsymbol{M}+\boldsymbol{W}\rVert_2\le\lVert\boldsymbol{M}\rVert_2+\lVert\boldsymbol{W}\rVert_2 \tag*{(46)}
∥ΔERIG∥2≤(Lf+12Lg∣Γ∣)∥η∥2⏟bound on ∥M∥2, from xref‘(39)‘+Lf2Lg∣Γ∣∥η∥22τ+O(∥η∥22)⏟bound on ∥W∥2, from xref‘(47)‘.(47)\lVert\Delta\mathrm{ERIG}\rVert_2\le\underbrace{\left(L_f+\tfrac{1}{2}L_g|\Gamma|\right)\lVert\boldsymbol{\eta}\rVert_2}_{\text{bound on }\lVert\boldsymbol{M}\rVert_2\text{, from }{xref}`(39)`}+\underbrace{\frac{L_f^2L_g|\Gamma|\lVert\boldsymbol{\eta}\rVert_2}{2\tau}+\mathcal{O}(\lVert\boldsymbol{\eta}\rVert_2^2)}_{\text{bound on }\lVert\boldsymbol{W}\rVert_2\text{, from }{xref}`(47)`}. \tag*{(47)}

Factoring out the model-response bound and defining R(τ)R(\tau) as the ratio of the two bounds,

∥ΔERIG∥2≤(Lf+12Lg∣Γ∣)∥η∥2⋅(1+R(τ)),R(τ)=Lf2Lg∣Γ∣2τ(Lf+12Lg∣Γ∣)+O(∥η∥2).(48)\lVert\Delta\mathrm{ERIG}\rVert_2\le\left(L_f+\tfrac{1}{2}L_g|\Gamma|\right)\lVert\boldsymbol{\eta}\rVert_2\cdot(1+R(\tau)),\qquad R(\tau)=\frac{L_f^2L_g|\Gamma|}{2\tau\left(L_f+\tfrac{1}{2}L_g|\Gamma|\right)}+\mathcal{O}(\lVert\boldsymbol{\eta}\rVert_2). \tag*{(48)}

Extension to Networks with ReLU Activations

Assumptions 1 and 2 hold directly for backbones with smooth activations, which in our evaluation are ViT-Base and DeiT3 with GELU, and EVA-02 and EfficientNet with SiLU. Such smooth activations are moreover the design norm in modern architectures, so the assumptions cover most recent models directly. For the remaining backbones the gradient is not continuous everywhere. ResNet101 is piecewise linear, with ∇f\nabla f piecewise constant on the network’s linear regions [32], while MobileNetV4 mixes ReLU stages with smooth attention blocks. In both cases ∇f\nabla f exists almost everywhere and Assumption 2 fails only on the kink boundaries.

The standard remedy is Gaussian smoothing, which convolves ff with the density φσ\varphi_\sigma of N(0,σ2I)\mathcal{N}(0,\sigma^2 I). The mollified network fσ=f∗φσf_\sigma= f * \varphi_\sigma is infinitely differentiable (fσ∈C∞f_\sigma\in C^\infty) and satisfies Assumptions 1 and 2 with Lf,σ≤LfL_{f,\sigma} \le L_f and Lg,σ=O(Lf/σ)L_{g,\sigma} = O(L_f/\sigma), respectively. Moreover, fσf_\sigma converges uniformly to ff as σ→0\sigma\to0; these are classical properties of Gaussian smoothing [21, 52].

Theorem 2 therefore applies verbatim to fσf_\sigma for every σ>0\sigma> 0, that is, to an arbitrarily close smooth surrogate of the network. The same device has already been seen in the XAI literature, and it underlies the analysis of explanation geometry of ReLU models [20]. Additionally, ∇fσ\nabla f_\sigma is exactly the gradient map that SmoothGrad [68] averages. Hence, our results extend to ReLU networks through this smoothing technique, with the decomposition stated for fσf_\sigma and its constants Lf,σL_{f,\sigma} and Lg,σL_{g,\sigma}.

Consistent with this, the stability gains in Table 3 are as large on the ReLU-based backbones as on the smooth ones, and the same holds for the adversarial stability results in Table 23 (Appendix A.15). Our empirical results are therefore not tied to the specific activation function used. Finally, as mentioned in § 3.2, exact Lipschitz constants of deep networks are intractable to compute, and we treat LfL_f and LgL_g only as analysis quantities rather than computed ones.

ModelMethodAS (↓)ALC (↓)AIS (↓)
ResNet101IG2.04000.0811117.301
ResNet101IDGI0.03340.001465.461
ResNet101ERIG (ours)1.3e−045.7e−061.033
EfficientNetIG1.26950.033165.008
EfficientNetIDGI0.02547.4e−0477.782
EfficientNetERIG (ours)9.0e−042.8e−0514.239
MobileNetV4IG1.76490.058340.621
MobileNetV4IDGI0.00924.8e−0439.747
MobileNetV4ERIG (ours)9.1e−043.6e−0511.086
ViT-BaseIG1.29320.028852.296
ViT-BaseIDGI0.03699.0e−0458.266
ViT-BaseERIG (ours)3.2e−048.0e−064.352
EVA-02IG1.16630.023765.522
EVA-02IDGI0.02956.2e−0465.661
EVA-02ERIG (ours)2.3e−044.9e−064.074
DeiT3IG1.39260.032377.716
DeiT3IDGI0.03298.0e−0461.983
DeiT3ERIG (ours)1.4e−043.4e−061.903

Table 23. ImageNet Adversarial Stability: ERIG is lowest on every backbone and metric, with a ∼90×\sim90 \times improvement over the next-best method on average. Moreover, its adversarial AS and ALC stay below every baseline’s uniform MS and LLC respectively.

Per-step Runtime Distributions

All runtime measurements are collected on 1000 validation samples from the ImageNet dataset and are measured on an NVIDIA A100-SXM4-40GB GPU using PyTorch and the ResNet101 model, with a batch size of 64. SmoothGrad is excluded here because its ensemble-based design requires NN independent attribution passes per example, making it extremely slow and its runtime not directly comparable on a per-step basis with path-based methods. Table 5 summarizes the per-step statistics for these methods and Figure 16 visualizes them with 5%–95% inter-percentile intervals; the shaded regions for IDGI and ERIG roughly align with each other, confirming that their per-step costs are approximately indistinguishable in practice.

MethodMean (ms/step)Std (ms/step)
BIG9.3720.843
GIG31.6820.473
IDGI1.8941.083
IG1.5860.671
ERIG (ours)1.9831.075

Table 5. Runtime: ERIG costs ∼\sim25% more per step than IG, with the softmax adding only ∼\sim5% over IDGI.

Figure 16. Per-step runtime with 5%–95% percentile ranges shaded in color. The vertical capped lines indicate the full range across samples. The horizontal lines inside the shaded regions represent the mean values.

Particularly, ERIG adds a small overhead relative to IG with its mean per-step runtime (1.983 ms) being ∼\sim0.4 ms (∼\sim25%) slower than IG (1.586 ms). Additionally, IDGI (1.894 ms) incurs a nearly identical overhead, indicating that the dominant additional cost shared by both methods is the element-wise squaring of the gradient vector, while the softmax normalization that distinguishes ERIG from IDGI adds only ∼\sim0.09 ms (∼\sim5%) per step. This marginal cost yields substantial gains in attribution stability, localization, and noise characteristics as demonstrated in the main text. All three methods pre-compute the full set of interpolated path points {x(k)}k=0K\{x^{(k)}\}_{k=0}^{K} before integration and evaluate them in batched forward-backward passes, yielding linear scaling with the step count (Figure 17, left).

Figure 17. Mean total runtime as a function of the number of integration steps over 1000 examples. The left subplot shows IG, IDGI, and ERIG, which scale linearly and overlap considerably. The right subplot shows BIG and GIG on a separate scale due to their higher costs.

BIG is, on average, ∼6×\sim6\times slower per step than IG because each step requires a Gaussian blur of the input, which involves a large spatial convolution in spite of batching. On the other hand, GIG is the slowest method at ∼20×\sim20\times the per-step cost of IG because its path is constructed iteratively with each intermediate point x(k)x^{(k)} depending on the computation at x(k−1)x^{(k-1)}. This sequential dependency prevents the path points from being pre-computed and batched together, forcing GIG into KK separate forward-backward passes, and making it fundamentally more expensive regardless of hardware (Figure 17, right).

ERIG Algorithm

The following algorithm provides a practical way of estimating ERIG attributions approximately:

Algorithm 1 Computing ERIG Attributions

  1. function ERIG⁡(f,Γ,τ,i)\operatorname{ERIG}(f, \Gamma, \tau, i)

  2. ERIG⁡i←0\operatorname{ERIG}_{i} \leftarrow0

  3. for x(k)∈Γx^{(k)} \in\Gamma do

  4. δf←f(x(k))−f(x(k−1))\delta f \leftarrow f(x^{(k)}) - f(x^{(k-1)})

  5. ERIG⁡i←ERIG⁡i+wi∗(g(x(k));τ)×δf\operatorname{ERIG}_{i} \leftarrow\operatorname{ERIG}_{i} + w_i^{*}(g(x^{(k)}); \tau) \times\delta f

  6. end for

  7. return ERIG⁡i\operatorname{ERIG}_{i}

  8. end function

SmoothGrad Ensembles

To evaluate whether the intrinsic properties of ERIG complement external stochastic smoothing techniques, we construct ensembles of each attribution method with SmoothGrad. To ensure a principled comparison with the vanilla methods discussed in the main text, we enforce a fixed computational budget of 200 total forward and backward passes. Specifically, we utilize N=10N = 10 noise samples with S=20S = 20 integration steps per sample (10×20=20010 \times20 = 200). While Smilkov et al. [68] suggest diminishing returns beyond approximately 50 noise samples for standalone SmoothGrad, path-based methods in an ensemble setting require a sufficiently high step count (≥20\ge20) to maintain a reliable Riemann sum approximation. This configuration represents a practical equilibrium between smoothing variance reduction and integration accuracy under strict resource constraints.

We maintain a fixed temperature of τ=3\tau= 3 for ERIG throughout these evaluations, consistent with the stability and noise experiments in the main text. We acknowledge that since SmoothGrad introduces external Gaussian noise which interacts with ERIG’s temperature-controlled weighting, the optimal temperature (τ\tau) for ensembles may deviate from the vanilla setting. A joint optimization over the hyper-parameter space of (τ\tau, number of samples, integration steps) is deferred to future work. Nonetheless, even with this untuned τ\tau, the ERIG ensemble achieves superior stability and noise scores across nearly all architectures (Table 6–Table 7). These results demonstrate that ERIG’s intrinsic stability and noise reduction mechanisms are orthogonal to external smoothing, confirming its effectiveness across diverse computational regimes.

ModelMethodMS (↓)LLC (↓)RIS (↓)
ResNet101SG + BIG0.31160.0139134.050
ResNet101SG + GIG0.26480.011828.502
ResNet101SG + IDGI0.00041.8e −05138.650
ResNet101SG + IG0.28010.0125141.100
ResNet101SG + ERIG (ours)0.00031.5e −054.029
EfficientNetSG + BIG0.03930.001212.868
EfficientNetSG + GIG0.10120.003227.078
EfficientNetSG + IDGI0.00262.1e −0514.696
EfficientNetSG + IG0.21790.006822.584
EfficientNetSG + ERIG (ours)8.2e −052.5e −062.003
MobileNetV4SG + BIG0.19780.007716.359
MobileNetV4SG + GIG0.18120.007131.392
MobileNetV4SG + IDGI0.00259.6e −0515.627
MobileNetV4SG + IG0.41410.016225.626
MobileNetV4SG + ERIG (ours)2.0e −046.5e −063.081
ViT-BaseSG + BIG0.06580.001710.109
ViT-BaseSG + GIG0.07150.001921.334
ViT-BaseSG + IDGI0.00420.000115.440
ViT-BaseSG + IG0.26620.006917.865
ViT-BaseSG + ERIG (ours)0.00235.9e −0537.984
EVA-02SG + BIG0.04910.00118.306
EVA-02SG + GIG0.04760.001120.135
EVA-02SG + IDGI0.00449.8e −0525.772
EVA-02SG + IG0.21830.004921.286
EVA-02SG + ERIG (ours)4.1e −059.2e −071.134
DeiT3SG + BIG0.11530.003010.650
DeiT3SG + GIG0.05610.001533.567
DeiT3SG + IDGI0.01480.000426.624
DeiT3SG + IG0.18130.004730.512
DeiT3SG + ERIG (ours)3.4e −058.9e −070.730

Table 6. SmoothGrad Stability: The ERIG ensemble improves stability by ∼67× on average over the next-best ensemble.

ModelMethodLINV-C (↑)LINV-P (↑)
ResNet101SG + BIG8.08047.9269
ResNet101SG + GIG8.22628.2521
ResNet101SG + IDGI13.916913.8936
ResNet101SG + IG8.27078.1122
ResNet101SG + ERIG (ours)18.818218.6851
EfficientNetSG + BIG8.28248.4709
EfficientNetSG + GIG8.25178.1945
EfficientNetSG + IDGI11.075411.0537
EfficientNetSG + IG7.47867.4437
EfficientNetSG + ERIG (ours)22.402722.3890
MobileNetV4SG + BIG6.65096.6404
MobileNetV4SG + GIG7.52277.3665
MobileNetV4SG + IDGI11.029310.9747
MobileNetV4SG + IG6.73446.5129
MobileNetV4SG + ERIG (ours)20.232920.1801
ViT-BaseSG + BIG8.47898.4251
ViT-BaseSG + GIG8.91958.6938
ViT-BaseSG + IDGI11.074410.9126
ViT-BaseSG + IG7.75177.4866
ViT-BaseSG + ERIG (ours)21.659021.4986
EVA-02SG + BIG8.78188.7877
EVA-02SG + GIG9.39689.1246
EVA-02SG + IDGI11.479011.4074
EVA-02SG + IG8.21568.0836
EVA-02SG + ERIG (ours)22.722022.6599
DeiT3SG + BIG7.86338.2038
DeiT3SG + GIG9.28749.2294
DeiT3SG + IDGI10.413810.3527
DeiT3SG + IG8.21408.2911
DeiT3SG + ERIG (ours)22.164322.1306

Table 7. SmoothGrad Noise: The ERIG ensemble attains the highest LINV in every cell, exceeding second by 4.8 to 11.8 units.

We are also interested in how the optimal τ\tau is affected when SmoothGrad’s external smoothing interacts with ERIG’s intrinsic smoothing. To this end, we sweep the ensemble temperature over τ∈[10−3,103]\tau\in[10^{-3}, 10^3], using 50 noise samples and 20 path steps per sample, on MobileNetV4, ResNet101, and ViT-Base. Figure 18 reports the four metrics as Figure 5, and the resulting pattern closely mirrors that observed in the latter. Fidelity remains on a broad plateau around the default value, localization peaks at an intermediate τ∈[0.1,3]\tau\in[0.1, 3], stability likewise reaches its minimum around these values, and noise reduction improves monotonically with τ\tau. The default τ=3\tau= 3 therefore remains a meaningful choice under ensembling as well.

Figure 18. Ablation of the softmax temperature τ\tau in SmoothGrad+ERIG ensembles across MobileNetV4, ResNet101, and ViT-Base models. All plots use a log-scale x-axis and a linear-scale y-axis, except stability, which also uses a log-scale y-axis.

Generalization Beyond ImageNet

To verify that our conclusions are not specific to the ImageNet family, we repeat the full evaluation protocol in two additional settings: the Oxford-IIIT Pet dataset [55] and the FunnyBirds benchmark [37]. This section provides the complete results behind the summary in § 4.6 of the main text.

Oxford-IIIT Pet

Oxford-IIIT Pet [55] is a 37-class fine-grained classification dataset that ships a pixel-aligned segmentation trimap with every image. Hence fidelity, stability, noise, and localization can all be evaluated within one non-ImageNet domain; for localization, the trimap’s boundary band is counted as background. For meaningful results, we evaluate the methods on models with good accuracy on that dataset. We therefore use five domain-fine-tuned public checkpoints spanning four architecture families, namely ResNet-50, MobileNetV2, ConvNeXt-B, ViT-B, and DINOv2-S. Table 8 to Table 11 report the results. ERIG remains competitive on fidelity, trailing the strongest baseline IDGI, on average, by only −0.40%, and outperforms the second-best method on localization by 20.2% on average across model–metric combinations. It additionally takes 24 of the 25 stability and noise cells, lowering MS and LLC by up to two orders of magnitude and posting LINV scores 6 to 8.7 units higher than those of the next-best method.

ModelMethodISC (↑)AIC (↑)DEL (↓)
ResNet50IG0.11310.38630.3668
ResNet50IDGI0.56490.82110.2497
ResNet50BIG0.11020.32650.3648
ResNet50GIG0.15790.40530.5246
ResNet50SmoothGrad0.10750.28680.6394
ResNet50ERIG (ours)0.57040.82280.2491
MobileNetV2IG0.12040.25200.3128
MobileNetV2IDGI0.58730.64640.2208
MobileNetV2BIG0.10270.19810.2836
MobileNetV2GIG0.16200.28070.3103
MobileNetV2SmoothGrad0.09180.15920.5581
MobileNetV2ERIG (ours)0.58980.63200.2194
ConvNeXt-BIG0.84340.88050.6499
ConvNeXt-BIDGI0.85880.90020.5568
ConvNeXt-BBIG0.77250.85210.6429
ConvNeXt-BGIG0.82110.87440.6724
ConvNeXt-BSmoothGrad0.40730.49820.8228
ConvNeXt-BERIG (ours)0.85770.90290.5562
ViT-BIG0.39190.62230.4886
ViT-BIDGI0.63490.74630.3680
ViT-BBIG0.29490.53320.5286
ViT-BGIG0.48760.64090.4345
ViT-BSmoothGrad0.16240.29200.6944
ViT-BERIG (ours)0.62220.73270.3808
DINOv2-SIG0.61930.72170.4638
DINOv2-SIDGI0.71040.80780.3783
DINOv2-SBIG0.43220.58790.4863
DINOv2-SGIG0.59140.74920.3766
DINOv2-SSmoothGrad0.20100.25180.6781
DINOv2-SERIG (ours)0.71500.80830.3766

Table 8. Table 8. Oxford-IIIT Pet Fidelity: ERIG is competitive and only trails the strongest baseline by 0.40% on average.

ModelMethodLINV-C (↑\uparrow)LINV-P (↑\uparrow)
ResNet50IG5.97805.8854
ResNet50IDGI9.61969.6044
ResNet50BIG6.81196.7880
ResNet50GIG7.42047.3881
ResNet50SmoothGrad7.88377.7788
ResNet50ERIG (ours)18.336518.3209
MobileNetV2IG5.76865.5927
MobileNetV2IDGI9.50979.4758
MobileNetV2BIG7.15617.1005
MobileNetV2GIG7.15727.1022
MobileNetV2SmoothGrad6.80956.5869
MobileNetV2ERIG (ours)17.235417.2044
ConvNeXt-BIG7.58207.5434
ConvNeXt-BIDGI9.33979.2465
ConvNeXt-BBIG7.30317.3255
ConvNeXt-BGIG7.98777.9592
ConvNeXt-BSmoothGrad10.973910.9233
ConvNeXt-BERIG (ours)17.008116.9278
ViT-BIG7.14277.1941
ViT-BIDGI10.274610.2591
ViT-BBIG7.74447.7675
ViT-BGIG7.36167.3296
ViT-BSmoothGrad8.64798.7419
ViT-BERIG (ours)17.981017.9928
DINOv2-SIG7.24197.0304
DINOv2-SIDGI9.49369.3007
DINOv2-SBIG7.35956.9305
DINOv2-SGIG7.33237.2198
DINOv2-SSmoothGrad8.84738.5305
DINOv2-SERIG (ours)18.202718.0135

Table 11. Oxford-IIIT Pet Noise: ERIG attains the highest LINV on every backbone, exceeding second by 6.0 to 8.7 units.

FunnyBirds

FunnyBirds [37] is a synthetic vision dataset purpose-built for part-based analysis of XAI methods. Every image is assembled from annotated bird parts, so the framework can map an attribution map to per-part importances and score it through part-level interventions. We evaluate ERIG and all compared path methods under its standard protocol, which evaluates along the completeness, correctness, and contrastivity dimensions, using the three checkpoints provided by the framework, namely ResNet-50, VGG16, and ViT-B. All metrics lie in [0,1][0, 1] and higher is better. Following [37], let xmx_m denote the mm-th sample of MM test images with label cmc_m and let f(xm)f(x_m) denote the model prediction. The part importance PIp(xm)\mathrm{PI}_p(x_m) of a bird part pp is the sum of the attribution values inside the part’s segmentation mask, dilated by a 5×55 \times5 kernel to include the part edges, and the important-part set is P(xm)={p∣PIp(xm)>t S(xm)}\mathrm{P}(x_m) = \{p \mid\mathrm{PI}_p(x_m) > t\,S(x_m)\}, where S(xm)S(x_m) is the total attribution of the map and the protocol sweeps the threshold tt over [0.01,0.50][0.01, 0.50], selecting it automatically to optimally trade off the comprehensiveness checks against distractibility. We explain the metrics below:

  • Controlled synthetic data check (CSDC, completeness). With Pcm,i′\mathcal{P}'_{c_m,i} the part subsets known to suffice for class cmc_m,

CSDC=1M∑m=1Mmax⁡i∣P(xm)∩Pcm,i′∣∣Pcm,i′∣(49)\mathrm{CSDC} = \frac{1}{M}\sum_{m=1}^{M} \max_i \frac{\left|\mathrm{P}(x_m) \cap\mathcal{P}'_{c_m,i}\right|}{\left|\mathcal{P}'_{c_m,i}\right|} \tag*{(49)}

measures how completely the explanation recovers a part set that suffices for the classification.

  • Preservation check (PC, completeness). With xm′x'_m the image keeping only the parts in P(xm)\mathrm{P}(x_m) and 1[⋅]\mathbb{1}[\cdot] the indicator function,

PC=1M∑m=1M1[f(xm′)=f(xm)](50)\mathrm{PC} = \frac{1}{M}\sum_{m=1}^{M} \mathbb{1}\left[f(x'_m) = f(x_m)\right] \tag*{(50)}

checks that the parts deemed important preserve the prediction.

ModelMethodIoU (↑)PA (↑)PF1 (↑)
ResNet50IG0.08290.61310.1348
ResNet50IDGI0.02000.59380.0361
ResNet50BIG0.05310.60300.0901
ResNet50GIG0.02510.59490.0442
ResNet50SmoothGrad0.11220.62320.1760
ResNet50ERIG (ours)0.15340.63400.2509
MobileNetV2IG0.07530.57000.1245
MobileNetV2IDGI0.02180.55360.0393
MobileNetV2BIG0.04750.56000.0819
MobileNetV2GIG0.03940.55560.0666
MobileNetV2SmoothGrad0.09930.57400.1602
MobileNetV2ERIG (ours)0.15420.59620.2507
ConvNeXt-BIG0.05600.56070.0941
ConvNeXt-BIDGI0.00880.54760.0166
ConvNeXt-BBIG0.04500.55470.0776
ConvNeXt-BGIG0.07450.54740.1243
ConvNeXt-BSmoothGrad0.10570.56620.1661
ConvNeXt-BERIG (ours)0.10330.57130.1773
ViT-BIG0.08280.69860.1370
ViT-BIDGI0.01690.69230.0309
ViT-BBIG0.05920.69550.1009
ViT-BGIG0.05020.69530.0862
ViT-BSmoothGrad0.12870.71410.2013
ViT-BERIG (ours)0.14050.70250.2335
DINOv2-SIG0.04770.55180.0820
DINOv2-SIDGI0.00720.54120.0138
DINOv2-SBIG0.02730.54480.0493
DINOv2-SGIG0.02660.54530.0477
DINOv2-SSmoothGrad0.07160.55890.1193
DINOv2-SERIG (ours)0.09730.56060.1690

Table 9. Table 9. Oxford-IIIT Pet Localization: ERIG improves localization by 20.2% on average over the second-best method.

  • Deletion check (DC, completeness). With xm′′x^{\prime\prime}_m the image with the parts in P(xm)\mathrm{P}(x_m) removed,

DC=1M∑m=1M1[f(xm′′)≠f(xm)](51)\mathrm{DC} = \frac{1}{M}\sum_{m=1}^{M} \mathds{1}\left[f(x^{\prime\prime}_m) \ne f(x_m)\right] \tag*{(51)}

checks that removing those parts changes the prediction.

  • Distractibility (Dist, completeness). With Pm′′\mathrm{P}^{\prime\prime}_m the parts whose individual removal changes the target logit by less than 5%,

Dist=1−1M∑m=1M∣P(xm)∩Pm′′∣∣Pm′′∣(52)\mathrm{Dist} = 1-\frac{1}{M}\sum_{m=1}^{M}\frac{\left|\mathrm{P}(x_m)\cap\mathrm{P}^{\prime\prime}_m\right|}{\left|\mathrm{P}^{\prime\prime}_m\right|} \tag*{(52)}

penalizes attribution mass placed on parts the model demonstrably ignores.

  • Single deletion (SD, correctness). With ρm\rho_m the Spearman correlation between the explanation’s part importances and the target-logit drops caused by deleting each part individually,

SD=12+12M∑m=1Mρm(53)\mathrm{SD} = \frac{1}{2}+\frac{1}{2M}\sum_{m=1}^{M}\rho_m \tag*{(53)}

scores whether the explanation ranks parts the way the model actually reacts to their removal.

  • Target sensitivity (TS, contrastivity). For two alternative target classes c^1\hat{c}_1 and c^2\hat{c}_2 that each share two distinct parts with the image, with PI′\mathrm{PI}^{\prime} and PI′′\mathrm{PI}^{\prime\prime} the summed importances on the parts specific to c^1\hat{c}_1 and c^2\hat{c}_2 respectively,

TS=12M∑m=1M(1[PI′(xm,c^1)>PI′(xm,c^2)]+1[PI′′(xm,c^1)<PI′′(xm,c^2)])(54)\mathrm{TS} = \frac{1}{2M}\sum_{m=1}^{M}\left(\mathds{1}\left[\mathrm{PI}^{\prime}(x_m,\hat{c}_1)>\mathrm{PI}^{\prime}(x_m,\hat{c}_2)\right]+\mathds{1}\left[\mathrm{PI}^{\prime\prime}(x_m,\hat{c}_1)<\mathrm{PI}^{\prime\prime}(x_m,\hat{c}_2)\right]\right) \tag*{(54)}

checks that the explanation shifts its mass to the class-specific parts when the explained target changes.

ModelMethodMS (↓\downarrow)LLC (↓\downarrow)RIS (↓\downarrow)
ResNet50IG0.60250.026923.753
ResNet50IDGI0.01155.1e−0424.485
ResNet50BIG0.10940.004913.453
ResNet50GIG0.19330.008637.823
ResNet50SmoothGrad0.08630.003921.475
ResNet50ERIG (ours)1.3e−045.8e−061.597
MobileNetV2IG1.55380.069416.533
MobileNetV2IDGI0.01215.4e−0410.858
MobileNetV2BIG0.17960.008014.231
MobileNetV2GIG0.31330.014022.746
MobileNetV2SmoothGrad0.33530.015011.542
MobileNetV2ERIG (ours)4.3e−041.9e−051.838
ConvNeXt-BIG0.26380.011811.672
ConvNeXt-BIDGI0.00492.2e−046.387
ConvNeXt-BBIG0.04120.00184.154
ConvNeXt-BGIG0.13340.006045.412
ConvNeXt-BSmoothGrad0.03030.001413.726
ConvNeXt-BERIG (ours)6.0e−052.7e−060.546
ViT-BIG0.60600.027010.238
ViT-BIDGI0.01050.000511.206
ViT-BBIG0.16300.00739.854
ViT-BGIG0.32750.014622.302
ViT-BSmoothGrad0.20840.009311.156
ViT-BERIG (ours)0.00640.000325.986
DINOv2-SIG0.16690.00756.395
DINOv2-SIDGI0.00351.6e−045.402
DINOv2-SBIG0.03800.00173.749
DINOv2-SGIG0.25700.011551.372
DINOv2-SSmoothGrad0.05060.002312.857
DINOv2-SERIG (ours)5.2e−052.3e−060.574

Table 10. Oxford-IIIT Pet Stability: ERIG improves stability by ∼38×\sim38\times on average over the next-best method.

In Table 12, Com. is the protocol’s completeness aggregate, ((CSDC+PC+DC)/3+Dist)/2((\mathrm{CSDC}+\mathrm{PC}+\mathrm{DC})/3+\mathrm{Dist})/2. ERIG leads the completeness aggregate by 3.4% and contrastivity by 2.1% on average across the three backbones while staying within 1.5% of the strongest baseline on correctness. It places first or second in all 21 protocol cells, and the cells it does not win lie within 1.8% of the best method. Under these definitions, ERIG’s wins concentrate where the metrics score the placement of attribution mass, covering the sufficient parts (CSDC) while avoiding the ignored ones (Dist). On the intervention and ranking checks (PC, DC, SD) it stays within 2% of the strongest baseline, with IDGI again the strongest on PC and DC, mirroring our ImageNet results.

ModelMethodCSDCPCDCDistCom.SDTS
Res-Net-50IG0.9200.9400.8800.8100.8620.5900.980
Res-Net-50IDGI0.9340.9580.9120.7880.8610.6150.958
Res-Net-50BIG0.7940.7000.6760.6990.7110.4400.491
Res-Net-50GIG0.7650.6840.6460.7990.7490.4370.780
Res-Net-50SmoothGrad0.8270.8360.8100.7950.8100.5450.873
Res-Net-50ERIG (ours)0.9680.9510.9050.8720.9070.6080.985
VGG16IG0.9200.9200.9200.9700.9450.6700.920
VGG16IDGI0.9450.9720.9580.9490.9540.6880.931
VGG16BIG0.7660.6680.6920.7230.7160.6120.582
VGG16GIG0.7700.6860.6940.8400.7780.5880.751
VGG16SmoothGrad0.9370.8200.8540.8240.8470.7090.903
VGG16ERIG (ours)0.9740.9640.9490.9780.9700.6980.955
ViT-BIG0.8900.8600.8500.9000.8830.6500.910
ViT-BIDGI0.9280.9410.9170.8710.9000.6640.918
ViT-BBIG0.7610.6500.7080.7350.7210.6680.533
ViT-BGIG0.7380.6280.6980.8410.7640.6630.659
ViT-BSmoothGrad0.9370.8260.8800.8340.8570.7800.856
ViT-BERIG (ours)0.9610.9330.9090.9240.9290.7660.948

Table 12. FunnyBirds Protocol: ERIG places first or second in all 21 cells, leading completeness by 3.4% and contrastivity by 2.1%.

Fixed vs. Path-Dependent Temperature

Proposition 1 permits a per-point temperature τ(t)\tau(t), while Sec. 4.1 adopts a constant τ\tau as a simplifying choice. Here, we further validate this choice empirically using a simple adaptive τ\tau schedule, demonstrating that designing an effective adaptive τ\tau is itself a nontrivial problem. We compare the constant τ=3\tau= 3 against ERIG with an adaptive temperature (AERIG), in which τ(t)\tau(t) decreases log-linearly from 101010^{10} to 10−1010^{-10} along the path. This schedule is a deliberately simple monotone choice that directly implements the premise that gradients near the baseline are less informative. The temperature remains high near the baseline, suppressing the influence of these gradients toward uniform weights, and decreases toward the input, concentrating the weights on the more informative gradients there. Table 13 to Table 16 report the ImageNet evaluation suite from the main paper across all six backbones. The schedule’s gains are scattered and small, whereas its losses are large and systematic. DEL worsens by 0.20 to 0.35 on every backbone. MS and LLC degrade by two to three orders of magnitude. LINV drops by 8.7 to 10.6 on five of the six backbones, and localization drops sharply for all backbones except EfficientNet. No backbone improves overall. Moreover, a monotone schedule is blind to changes in the model output, reaching its smallest τ\tau in the saturation plateau, where concentrating the weights amplifies noise rather than signal (Figure 2). The constant τ\tau is therefore also justified empirically. We acknowledge that this single schedule does not constitute an exhaustive exploration of adaptive temperatures. A schedule that tracks changes in the model output would itself constitute an academic contribution, much as refinements of the Riemann approximation [66], the integration path [43, 92], and the baseline [23, 47, 65, 74, 97] have each constituted contributions to the IG family.

ModelMethodISC (↑)AIC (↑)DEL (↓)
ViT-BaseERIG0.69300.90130.4902
AERIG0.45250.78060.7405
EVA02ERIG0.59130.89070.5373
AERIG0.48880.85880.7786
MobileNetV4ERIG0.41180.72340.3626
AERIG0.38590.65010.6959
EfficientNetERIG0.45210.83230.4562
AERIG0.47670.77920.6889
ResNet101ERIG0.48930.66100.3270
AERIG0.53530.67930.6730
DeiT3ERIG0.73200.90100.5561
AERIG0.60940.86410.7512

Table 13. AERIG Ablation Fidelity: The adaptive schedule worsens DEL everywhere, winning only three of eighteen cells.

ModelMethodLINV-C (↑)LINV-P (↑)
ViT-BaseERIG19.34319.201
AERIG10.39010.382
EVA02ERIG20.12820.115
AERIG10.55210.572
MobileNetV4ERIG18.13018.113
AERIG9.32529.3624
EfficientNetERIG20.70520.712
AERIG10.06110.272
ResNet101ERIG16.80716.778
AERIG19.77219.746
DeiT3ERIG19.57819.584
AERIG10.79510.932

Table 16. AERIG Ablation Noise: The adaptive schedule loses 8.7 to 10.6 LINV on five of six backbones.

ModelMethodIoU (↑)PA (↑)PF1 (↑)
ViT-BaseERIG0.12030.70540.1893
AERIG0.05990.63820.0837
EVA02ERIG0.13170.70540.2031
AERIG0.03880.65050.0546
MobileNetV4ERIG0.16150.65210.2424
AERIG0.09880.56940.1340
EfficientNetERIG0.10850.59820.1714
AERIG0.12240.54750.1640
ResNet101ERIG0.16170.65800.2419
AERIG0.04230.61280.0588
DeiT3ERIG0.11790.66370.1819
AERIG0.01330.63680.0180

Table 14. AERIG Ablation Localization: The constant temperature is better in every cell except EfficientNet IoU.

ModelMethodMS (↓)LLC (↓)RIS (↓)
ViT-BaseERIG3.0e−049.0e−064.9460
AERIG0.02556.6e−0450.070
EVA02ERIG6.7e−051.5e−061.6010
AERIG0.01874.2e−0466.966
MobileNetV4ERIG1.0e−045.4e−062.1960
AERIG0.04730.001872.558
EfficientNetERIG6.7e−052.1e−061.2710
AERIG0.02658.3e−0462.522
ResNet101ERIG8.6e−053.8e−061.0660
AERIG0.06210.0028104.7
DeiT3ERIG3.9e−051.0e−060.7560
AERIG0.01574.1e−0483.403

Table 15. AERIG Ablation Stability: The adaptive schedule degrades MS and LLC by two to three orders of magnitude.

ModelBaselineIoU (↑\uparrow)PA (↑\uparrow)PF1 (↑\uparrow)
EVA-02zero0.13170.70540.2031
EVA-02ones0.12120.70460.1879
EVA-02random color0.12450.70600.1925
EVA-02channel mean0.12280.70450.1897
EVA-02channel median0.12300.70530.1902
EVA-02uniform0.09060.70320.1395
EVA-02Gaussian0.03190.69650.0507
EVA-02blur0.09730.70000.1516
EVA-02pixel perm.0.05230.69870.0814
EVA-02window perm.0.02700.69490.0428
EVA-02magnitude scr.0.04160.69780.0660
EVA-02phase scr.0.05760.69980.0905
ResNet101zero0.16170.65800.2419
ResNet101ones0.15680.65610.2361
ResNet101random color0.15590.65600.2347
ResNet101channel mean0.15560.65630.2346
ResNet101channel median0.15570.65640.2347
ResNet101uniform0.16600.65770.2471
ResNet101Gaussian0.15810.65500.2363
ResNet101blur0.15390.65620.2325
ResNet101pixel perm.0.16420.65730.2436
ResNet101window perm.0.13690.64350.2081
ResNet101magnitude scr.0.16280.65570.2422
ResNet101phase scr.0.16110.65560.2410

Table 18. Baseline Ablation Localization: Constant baselines stay within 8.4% on IoU and PF1; the rest are backbone-specific.

ModelBaselineMS (↓\downarrow)LLC (↓\downarrow)RIS (↓\downarrow)
EVA-02zero6.7e–051.5e–061.601
EVA-02ones6.2e–051.4e–061.410
EVA-02random color6.4e–051.4e–061.465
EVA-02channel mean6.2e–051.4e–061.413
EVA-02channel median6.2e–051.4e–061.411
EVA-02uniform6.4e–051.4e–061.445
EVA-02Gaussian6.2e–051.4e–061.410
EVA-02blur6.2e–051.4e–061.439
EVA-02pixel perm.6.2e–051.4e–061.410
EVA-02window perm.2.0e–045.2e–0647.530
EVA-02magnitude scr.1.0e–042.2e–0629.804
EVA-02phase scr.6.2e–051.4e–061.414
ResNet101zero8.6e–053.8e–061.066
ResNet101ones2.0e–041.0e–052.599
ResNet101random color1.1e–025.0e–0489.584
ResNet101channel mean2.0e–049.7e–062.741
ResNet101channel median2.0e–048.0e–062.340
ResNet101uniform1.0e–046.4e–062.046
ResNet101Gaussian1.0e–046.3e–062.036
ResNet101blur2.0e–049.4e–062.641
ResNet101pixel perm.1.0e–046.3e–062.040
ResNet101window perm.3.0e–041.5e–0549.918
ResNet101magnitude scr.2.0e–048.3e–062.670
ResNet101phase scr.1.0e–046.4e–062.048

Table 19. Baseline Ablation Stability: Effects are backbone-specific rather than systematic across the twelve baselines.

Baseline Sensitivity

As can be seen from Table 17 to Table 20, only one pattern is consistent across both backbones: the constant references closely track the zero default. ISC, AIC, DEL, and PA remain within 0.7% of the family mean on EVA-02 and within 4.0% on ResNet101, while IoU and PF1 remain within 8.4% on both backbones, indicating that ERIG’s fidelity and localization do not depend critically on the black baseline. Beyond this, the effects of the baseline are backbone-specific rather than systematic, and the same family can exhibit opposite trends across backbones. Uniform noise increases ISC by 4.5% while reducing IoU by 31% on EVA-02, yet decreases ISC by 7% and achieves the best IoU on ResNet101. Standard normal noise leaves localization on ResNet101 largely unchanged but reduces EVA-02’s IoU by three quarters. The per-image random color baseline is indistinguishable from the other constant baselines on EVA-02, yet degrades stability on ResNet101 by approximately two orders of magnitude. The structure-destroying replacements substantially impair localization on EVA-02, with IoU decreasing by 56 to 80%, whereas ResNet101 experiences a decrease of at most 15%. Even within the same family, the window permutation increases RIS thirtyfold on EVA-02, while the phase scramble produces results comparable to the zero default. Baseline choice therefore has model-specific effects that are not captured by family-level summaries. The consistent finding is that constant references closely reproduce the behavior of the default baseline.

ModelBaselineISC (↑\uparrow)AIC (↑\uparrow)DEL (↓\downarrow)
EVA-02zero0.59130.89070.5373
EVA-02ones0.58720.89230.5368
EVA-02random color0.58930.89720.5374
EVA-02channel mean0.59060.89640.5369
EVA-02channel median0.59010.89610.5381
EVA-02uniform0.61780.90200.5491
EVA-02Gaussian0.61180.90420.5577
EVA-02blur0.57980.91100.5638
EVA-02pixel perm.0.59840.89620.5487
EVA-02window perm.0.54900.91280.6587
EVA-02magnitude scr.0.60040.90130.5482
EVA-02phase scr.0.60310.90850.5568
ResNet101zero0.48930.66100.3270
ResNet101ones0.49170.68810.3254
ResNet101random color0.48170.67780.3276
ResNet101channel mean0.48990.68230.3275
ResNet101channel median0.48860.68190.3272
ResNet101uniform0.45520.66340.3293
ResNet101Gaussian0.45650.65500.3304
ResNet101blur0.47750.67150.3373
ResNet101pixel perm.0.44700.65270.3273
ResNet101window perm.0.42450.66050.3759
ResNet101magnitude scr.0.44460.65860.3285
ResNet101phase scr.0.44120.64490.3412

Table 17. Baseline Ablation Fidelity: The constant baselines track the zero default within 0.7–4.0%.

ModelBaselineLINV-C (↑\uparrow)LINV-P (↑\uparrow)
EVA-02zero20.128520.1150
EVA-02ones20.090220.0854
EVA-02random color20.312720.2931
EVA-02channel mean20.200020.1827
EVA-02channel median20.289220.2729
EVA-02uniform22.394222.2114
EVA-02Gaussian23.852123.6896
EVA-02blur21.431321.3799
EVA-02pixel perm.23.631923.6009
EVA-02window perm.24.615624.4948
EVA-02magnitude scr.23.793423.7436
EVA-02phase scr.22.998522.9185
ResNet101zero16.806916.7781
ResNet101ones16.236116.2020
ResNet101random color16.363416.3341
ResNet101channel mean16.226416.1936
ResNet101channel median16.232416.2009
ResNet101uniform18.519318.4302
ResNet101Gaussian19.595619.5321
ResNet101blur16.706916.6935
ResNet101pixel perm.19.337219.2665
ResNet101window perm.18.560218.5158
ResNet101magnitude scr.19.638919.5709
ResNet101phase scr.18.316418.3061

Table 20. Baseline Ablation Noise: LINV stays within 4.5 units of the zero default across all twelve baselines.

Validation of the Noise Metrics

To relate LINV to existing complexity, compressibility, and sparsity measures, we evaluate five such measures alongside LINV on the ImageNet evaluation of the main paper (six methods on six backbones). Each of the new metrics quantifies how strongly attribution mass concentrates, so we refer to them collectively as concentration measures. Since these measures originate in digital-image analysis, they are computed on the 8-bit grayscale rendering u∈{0,…,255}Nu \in\{0,\ldots,255\}^{N} of the channel-summed attribution map, except the Gini coefficient, which uses the attribution visualization vv.

  • Histogram entropy. Hhist=−∑ℓ=0255qℓlog⁡2qℓ∈[0,8]H_{\mathrm{hist}} = -\sum_{\ell=0}^{255} q_{\ell}\log_{2}q_{\ell} \in[0,8] bits, with qℓq_{\ell} the fraction of pixels of uu at gray level ℓ\ell.

  • JPEG compressibility. JPEGratio\mathrm{JPEG}_{\mathrm{ratio}}, the encoded-to-raw byte ratio of uu at quality 90 (a practical stand-in for Kolmogorov complexity), and JPEGRMSE\mathrm{JPEG}_{\mathrm{RMSE}}, the root-mean-square error between uu and its JPEG reconstruction.

  • Gini sparsity [12]. GS=∑i=1N(2i−N−1)v(i)N∑i=1Nv(i)∈[0,1)\mathrm{GS} = \frac{\sum_{i=1}^{N}(2i-N-1)v_{(i)}}{N\sum_{i=1}^{N}v_{(i)}} \in[0,1) over the sorted entries v(1)≤⋯≤v(N)v_{(1)} \le\cdots\le v_{(N)} of vv; higher means attribution mass concentrated on fewer features.

  • Entropy complexity [9]. H=−∑i=1Npiln⁡piln⁡N∈[0,1]H = -\frac{\sum_{i=1}^{N}p_i\ln p_i}{\ln N} \in[0,1] with pi=ui/∑jujp_i = u_i/\sum_j u_j, the normalized Shannon entropy of the fractional-contribution distribution; a value of 1 means maximally spread.

Two observations follow from Table 21. First, the concentration measures disagree with each other. By the four measures computed on uu, ERIG produces the most concentrated and most compressible maps, with the lowest HH, HhistH_{\mathrm{hist}}, and JPEGRMSE\mathrm{JPEG}_{\mathrm{RMSE}} on all six backbones, while the Gini of the visualization puts IDGI first on five of six backbones with ERIG in mid-field. IDGI holds its mass in a few isolated, extremely large values, which the Gini coefficient rewards, yet around those spikes its maps still flicker with fine-grained variation, so as images they remain busy; ERIG’s maps have no towering pixels but a genuinely quiet background, which the entropy and compressibility measures reward. Second, no concentration measure orders the methods the way noise does. SmoothGrad is the most diffuse baseline according to histogram entropy and both JPEG measures, hence the least concentrated and, by those measures’ logic, the noisiest, yet cleaner than IG, BIG, and GIG by LINV on every backbone; GIG is more concentrated than SmoothGrad by those proxies yet noisier by LINV on every backbone; and IDGI and ERIG nearly coincide on the entropy and JPEG measures while LINV nearly doubles between them.

ModelMethodLINV-CLINV-PHHHhistH_{\mathrm{hist}}JPEGratio_{\mathrm{ratio}}JPEGRMSE_{\mathrm{RMSE}}GS
ResNet101IG6.5876.4890.99765.9370.39573.3450.6989
ResNet101IDGI9.4319.4020.89104.5280.27902.6730.8346
ResNet101BIG6.6496.6050.99785.6740.38683.3250.7304
ResNet101GIG7.9657.9090.99804.6000.32722.9980.8036
ResNet101SmoothGrad10.1359.9180.99686.6750.49253.7050.6000
ResNet101ERIG (ours)16.80716.7780.87484.2290.26732.5730.7467
EfficientNetIG6.8776.8570.99756.3180.43333.5320.6552
EfficientNetIDGI10.24310.2460.92084.8030.32222.9680.8122
EfficientNetBIG7.9178.0540.99776.0080.38073.2890.6974
EfficientNetGIG7.3647.2320.99814.6430.31972.9830.7984
EfficientNetSmoothGrad8.7298.7100.99726.6050.46793.7050.6151
EfficientNetERIG (ours)20.70520.7120.89594.4580.32122.8670.7315
MobileNetV4IG5.9715.7930.99726.4060.46203.6480.6490
MobileNetV4IDGI10.08610.0730.93895.0710.33063.0150.7940
MobileNetV4BIG6.5066.4760.99775.9580.40353.4320.7078
MobileNetV4GIG6.5986.4470.99794.6830.33393.0520.7980
MobileNetV4SmoothGrad7.8487.6510.99656.8370.52433.8020.5814
MobileNetV4ERIG (ours)18.13018.1130.89594.6320.30882.7950.7132
ViT-BaseIG7.2227.0610.99796.0140.39003.3330.7004
ViT-BaseIDGI10.16210.0080.92073.0870.15031.8590.9296
ViT-BaseBIG8.0367.9450.99795.9450.38303.2520.6968
ViT-BaseGIG8.0057.9180.99824.5460.30232.8890.8188
ViT-BaseSmoothGrad9.2608.8760.99766.3520.41853.4070.6483
ViT-BaseERIG (ours)19.34319.2010.82342.4910.12871.6480.8412
EVA-02IG7.5687.5230.99786.2910.41643.3970.6499
EVA-02IDGI10.80310.7780.94043.9670.21592.2700.8855
EVA-02BIG8.4678.4220.99815.9160.36953.2130.7029
EVA-02GIG8.5058.3320.99824.6990.32092.9560.7917
EVA-02SmoothGrad9.5959.1270.99726.7370.48663.6540.5784
EVA-02ERIG (ours)20.12820.1150.86423.3460.19102.0120.7977
DeiT3IG7.1587.5700.99825.6730.32722.9970.7478
DeiT3IDGI10.24710.2540.91773.4390.19072.0530.9131
DeiT3BIG7.7478.1020.99815.6210.31242.7720.7460
DeiT3GIG8.6588.7210.99844.4580.28912.8620.8273
DeiT3SmoothGrad9.5929.3610.99825.9710.36433.3000.6971
DeiT3ERIG (ours)19.57819.5840.85233.0110.18121.9160.8177

Table 21. LINV vs. Concentration: Compared measures disagree not only with LINV but also among themselves on the overall order.

Concentration can nevertheless serve as a footprint of noise removal. Small-amplitude components of gradient irregularities can propagate into the attribution with broad spatial support, assigning nonzero relevance to pixels that provide little or no evidence for the prediction. This broad support spreads the attribution values across gray levels, increasing both histogram and contribution entropies, introducing high-frequency content that is less amenable to compression, and shifting attribution mass away from relevant features. Suppressing these components has the opposite effect, reducing variation in the background and concentrating the attribution mass around the evidence. ERIG’s entropy-regularized weighting leaves exactly this footprint, the lowest entropies and the most compressible maps in Table 21, and its localization confirms the reading; with the best IoU and PF1 on all six backbones (Table 2), the mass that survives its weighting sits on the ground-truth object regions, so its concentration comes from removing off-object, noise-like relevance rather than from discarding evidence.

Concentration can, however, arise without noise removal. Because concentration measures are functionals of the value distribution alone, they provide no information about the spatial location of the attribution mass. An attribution map concentrated on the object and one concentrated on a small number of noise pixels can therefore exhibit similar concentration statistics, with their distinction captured instead by localization metrics. IDGI illustrates this limitation. It ranks second to ERIG according to HH on every backbone and first according to GS on five of six backbones, yet exhibits the weakest localization performance throughout. Its attribution mass is concentrated in a small number of high-amplitude gradient irregularities, which its direct squared-gradient coupling passes through without attenuation. Moreover, concentration measures can conflate noise with the spatial distribution of the underlying evidence by mistaking structure for unwanted variation, making their absolute values insufficient for assessing noise removal. Concentration is therefore best interpreted as a potential byproduct of denoising rather than as a direct measure of it.

The correlations quantify this lack of correspondence. Table 22 reports Spearman rank correlations between LINV, the five concentration measures, and the stability metrics from Table 3. We compute correlations across all 36 model-method cells and separately across the 30 cells excluding ERIG, thereby controlling for the possibility that a single method drives the observed correlations. Each measure is evaluated using its nominal orientation, with lower values receiving better ranks for HH, HhistH_{\mathrm{hist}}, JPEGratio_{\mathrm{ratio}}, JPEGRMSE_{\mathrm{RMSE}}, MS, LLC, and RIS, and higher values receiving better ranks for GS. Thus, a positive coefficient indicates agreement with LINV’s ordering in which higher values correspond to cleaner attributions. For the concentration measures, however, this orientation is a ranking convention rather than a claim about their relationship to attribution quality. LINV exhibits strong agreement with the stability metrics associated with gradient noise and irregularities. Its Spearman correlation with MS and LLC ranges from +0.91 to +0.93 and remains between +0.84 and +0.89 when ERIG is excluded, while its agreement with RIS is weaker, consistent with RIS being the only scale-normalized metric among the three which LINV is not. In contrast, LINV exhibits at most moderate agreement with any concentration measure (ρ≤0.66\rho\le0.66), which decreases to ρ≤0.43\rho\le0.43 when ERIG is excluded. Empirically, LINV therefore behaves more consistently with the stability family than with the concentration family.

CellsMetricHHHhistH_{\mathrm{hist}}JPEGratio_{\mathrm{ratio}}JPEGRMSE_{\mathrm{RMSE}}GSMSLLCRIS
All 36 cellsLINV-C+0.66+0.55+0.57+0.60+0.36+0.91+0.92+0.52
All 36 cellsLINV-P+0.65+0.56+0.60+0.63+0.38+0.92+0.93+0.55
Without ERIGLINV-C+0.43+0.28+0.36+0.38+0.27+0.84+0.86+0.17
Without ERIGLINV-P+0.42+0.31+0.41+0.42+0.30+0.87+0.89+0.23

Table 22. LINV Correlations: LINV correlates at +0.91 to +0.93 with the stability metrics but at most +0.66 with any other compared measure. Even with ERIG excluded, the stability correlations stay at +0.86 to +0.89 while the rest collapse further.

None of the concentration measures directly quantifies noise. HH, HhistH_{\mathrm{hist}}, and GS are functionals of the value distribution alone. Spatially scrambling an attribution map, thereby transforming its structure into salt-and-pepper noise, leaves these measures exactly unchanged, whereas LINV, which is spatial by construction, identifies the scrambled map as highly noisy through its Laplacian-like convolution (A∗S)(A * S) and patch statistics. The JPEG measures quantify overall fine-scale content and cannot distinguish unstructured fluctuations from legitimate fine-grained evidence. For example, an attribution map with sharp object boundaries may compress less effectively than a blurred map without necessarily being noisier. In contrast, LINV-C and LINV-P target the quantity of interest directly, namely the variance of the spatially unstructured component, using the Laplacian-like convolution of [39] and the patch-PCA eigenvalue tail of [16], respectively. The concentration measures therefore serve as secondary descriptive statistics, whereas LINV directly quantifies the noise component of attribution evaluation. The two LINV variants also exhibit close agreement across all backbones (Table 4), indicating that either variant can serve as a standalone noise metric.

Adversarial Stability

The uniform stability protocol of § 4.4 samples isotropic perturbations, and no finite Monte Carlo sample can cover every direction of a high-dimensional input space. We therefore complement it with adversarially selected perturbation directions, computed by projected gradient ascent in the spirit of the attribution attacks of Ghorbani et al. [28] and Dombrowski et al. [20]. At the same perturbation radius r=0.1r = 0.1 as the uniform protocol, the attack maximizes the dissimilarity between the original and perturbed attribution maps while keeping the prediction unchanged; the constraint is enforced by a margin penalty during the ascent and by a final feasibility check afterwards. With ex\mathbf{e}_{x} the attribution map at input xx, ff the model output, yy the predicted class, and [z]+=max⁡(z,0)[z]_{+} = \max(z, 0), each restart ascends

J(δ)=1−⟨ex+δ, ex⟩∥ex+δ∥2 ∥ex∥2−λ[max⁡c≠yfc(x+δ)−fy(x+δ)+m]+(55)\mathcal{J}(\boldsymbol{\delta}) = 1 - \frac{\left\langle\mathbf{e}_{\mathbf{x}+\boldsymbol{\delta}},\, \mathbf{e}_{\mathbf{x}} \right\rangle}{\lVert\mathbf{e}_{\mathbf{x}+\boldsymbol{\delta}} \rVert_{2}\,\lVert\mathbf{e}_{\mathbf{x}} \rVert_{2}} - \lambda\left[\max_{c \ne y} f_c(\mathbf{x}+\boldsymbol{\delta}) - f_y(\mathbf{x}+\boldsymbol{\delta}) + m\right]_{+} \tag*{(55)}

with sign-gradient steps under a cosine-decayed step size ηt\eta_t,

δt+1=Πr(δt+ηtsign⁡(∇δJ(δt))),ηt=η0r(0.55+0.45cos⁡πtT),(56)\boldsymbol{\delta}_{t+1} = \Pi_r\left(\boldsymbol{\delta}_t + \eta_t \operatorname{sign}\left(\nabla_{\boldsymbol{\delta}}\mathcal{J}(\boldsymbol{\delta}_t)\right)\right), \qquad\eta_t = \eta_0 r\left(0.55 + 0.45\cos\frac{\pi t}{T}\right), \tag*{(56)}

where Πr\Pi_r projects back onto the radius-rr perturbation set. The defaults are λ=1\lambda= 1 and m=0.0m = 0.0, with mm the logit margin the perturbation must maintain between the predicted class’s logit and the next largest. The ascent is run for T=100T = 100 steps from each of R=3R = 3 independent random restarts per image; all three are drawn with the uniform protocol’s perturbation law, the first from a smaller ball of radius 0.01r0.01r so that the ascent also starts close to the input, the other two at the full radius rr. Instead of relying on a single ascent, the restarts therefore probe the non-convex objective from starting points near the unperturbed input as well as across the full perturbation set. The ascent needs second-order derivative information, since the attribution is built from input gradients and differentiating it with respect to δ\boldsymbol{\delta} passes through second derivatives of ff. ReLU activations have a second derivative that is not only undefined at their kinks, but also zero away from them. This degrades the optimization signal and can make the ascent direction vanish. Following [20], ReLU backbones are therefore softplus-smoothed during the attack optimization only, restoring nonzero curvature in the spirit of the smooth surrogates of Appendix A.7.3. Every reported quantity, however, is evaluated on the original network at full path resolution, so that the smoothing cannot affect any metric computation.

We report the adversarial counterparts of the three stability metrics of Table 3, Adversarial Sensitivity (AS), Adversarial Lipschitz Constant (ALC), and Adversarial Input Stability (AIS). They are the same functionals as MS, LLC, and RIS, with the worst case taken over adversarially chosen directions instead of uniform draws. Let A(x)={x+δ(k)}k=1R\mathcal{A}(x) = \left\{x + \delta^{(k)}\right\}_{k=1}^{R} collect the perturbed inputs the attack retains for image xx, i.e. the endpoint δ(k)\delta^{(k)} of each of the RR restarts after the feasibility backtracking that shrinks any perturbation still changing the predicted class. Then

AS⁡(x,r)=max⁡x′∈A(x)∥ex−ex′∥2(57)\operatorname{AS}(x,r) = \max_{x' \in\mathcal{A}(x)} \lVert\mathbf{e}_x - \mathbf{e}_{x'} \rVert_2 \tag*{(57)}
ALC⁡(x,r)=max⁡x′∈A(x)∥ex−ex′∥2∥x−x′∥2(58)\operatorname{ALC}(x,r) = \max_{x' \in\mathcal{A}(x)} \frac{\lVert\mathbf{e}_x - \mathbf{e}_{x'} \rVert_2}{\lVert x - x' \rVert_2} \tag*{(58)}
AIS⁡(x,r)=max⁡x′∈A(x)∥ex−ex′∥2/∥ex∥2∥x−x′∥2/∥x∥2(59)\operatorname{AIS}(x,r) = \max_{x' \in\mathcal{A}(x)} \frac{\lVert\mathbf{e}_x - \mathbf{e}_{x'} \rVert_2 / \lVert\mathbf{e}_x \rVert_2}{\lVert x - x' \rVert_2 / \lVert x \rVert_2} \tag*{(59)}

Every per-sample score is thus a maximum over all RR retained adversarial perturbations, not the outcome of a single attack run, and the dataset median of these per-sample worst cases is what Table 23 reports. The attack set consists of the methods whose attributions are differentiable functions of the input, namely IG, IDGI, and ERIG. The quantile-based feature selection, the saturation rule, and the repeat-until inner loop in Algorithm 1 of GIG [43] are input-dependent discrete operations that render the attribution piecewise-defined and discontinuous, so gradient-based worst-case optimization is ill-posed for it. Blur IG is excluded on cost rather than principle; the attack is already second order as noted above, and Blur IG’s path points are blurred copies of the input, which increases that cost considerably. We keep the uniform-perturbation results in § 4.4, as they follow the established protocols of the metrics’ original papers [4, 5, 94] and enable comparison with a wider set of methods and with prior work.

As Table 23 shows, the conclusion of Table 3 carries over unchanged to adversarially selected directions. ERIG attains the lowest score on every backbone and every metric, improving worst-case attribution stability by ∼90×\sim90 \times on average over the next-best method and reaching up to four orders of magnitude over IG. Notably, ERIG’s adversarial worst case remains below every baseline’s uniform MS and LLC in Table 3; even adversarially selected directions cannot push ERIG’s attribution shift up to the level the baselines already reach under isotropic noise, consistent with the direction-independent O(1/(2τ))\mathcal{O}(1/(2\tau)) weight-variation bound of Theorem 2.

Signed ERIG

In Eq. (4) the weights are nonnegative, but the sign of each contribution comes from the function increments dfdf, which are signed. ERIG attributions therefore can be negative; features weighted strongly on path segments where ff decreases receive negative mass. Squaring removes the sign of the per-feature gradient inside the weighting, not the sign of the attribution, and IDGI [93] shares exactly this structure. The attribution sign is nevertheless coarsened, and it is true that downstream applications such as absent-concept explanation [38] do exploit sign information. We therefore ablate a signed variant (SERIG) with weights w~i=wi∗⋅sign⁡(gi)\tilde{w}_i = w_i^* \cdot\operatorname{sign}(g_i), which restores the per-step gradient sign at no extra computational cost while preserving the magnitude profile (Figure 19). Tables 24 to 27 report the ImageNet evaluation suite of the main paper.

Schematic showing ERIG and SERIG per-step contributions and running sums for on-object and background features

Figure 19. Schematic – Why the Signed Variant Cancels: ERIG’s per-step terms carry the sign of δf(k)\delta f^{(k)}, which is shared by every feature, so nothing cancels if ff keeps increasing, and both on-object and background features accumulate alike. SERIG multiplies the weights by sign⁡(gi)\operatorname{sign}(g_i). This sign is usually stable on the object but alternating along the path for background, so only the background feature cancels toward zero in the aggregation (summation). Here, the hypothetical per-step weight magnitudes follow the movement of the model output along the path (Figure 2), small over the flat and saturated segments and large where the output changes rapidly.

The ablation reveals a consistent trade-off. Restoring the gradient sign substantially improves localization across all six backbones, with IoU increasing from the 0.11–0.16 range to the 0.18–0.25 range and PFI improving by 27–85%. This improvement comes at the cost of fidelity and stability. ISC decreases by 0.04–0.30, while MS and LLC worsen by factors of 2–7 on ViT-Base and EVA-02 and by one to two orders of magnitude on the remaining backbones, although SERIG remains substantially more stable than IG, BIG, and GIG in Table 3. The underlying mechanism is that the two maps capture different properties (Figure 19). ERIG accumulates the magnitude of each feature’s influence on the output along the integration path, irrespective of the direction of that influence, and thus produces a sensitivity map. The signed variant accumulates attribution only when a feature’s influence has a consistent direction, causing contributions with opposing signs along the path to cancel and producing a coherence map. This directional coherence is precisely the property favored by localization. Object regions tend to provide evidence that is more directionally consistent along the integration path, whereas background and texture can influence the prediction in more variable directions, as these regions vary substantially across images while the object remains comparatively consistent. Cancellation therefore suppresses precisely the off-object attribution mass. In contrast, fidelity probes favor sensitivity. Deletion and insertion measure whether removing attributed pixels changes the model output, and a directionally unstable pixel can still exert a positive influence on ff at the input xax^a. The signed map can therefore cancel pixels to which the model demonstrably responds, while the fidelity probes continue to detect output

ModelMethodISC (↑)AIC (↑)DEL (↓)
ViT-BaseERIG0.69300.90130.4902
SERIG0.41240.65230.7479
EVA02ERIG0.59130.89070.5373
SERIG0.42100.76210.8037
MobileNetV4ERIG0.41180.72340.3626
SERIG0.31050.50860.7007
EfficientNetERIG0.45210.83230.4562
SERIG0.41060.70300.7140
ResNet101ERIG0.48930.66100.3270
SERIG0.31120.43790.7003
DeiT3ERIG0.73200.90100.5561
SERIG0.43630.63110.7743

Table 24. SERIG Ablation Fidelity: ERIG beats the signed variant in every cell, by factors up to 1.7× on ISC and 2.1× on DEL.

ModelMethodIoU (↑)PA (↑)PF1 (↑)
ViT-BaseERIG0.12030.70540.1893
SERIG0.18420.51300.2647
EVA02ERIG0.13170.70540.2031
SERIG0.18980.52320.2728
MobileNetV4ERIG0.16150.65210.2424
SERIG0.21740.52750.3085
EfficientNetERIG0.10850.59820.1714
SERIG0.22650.50970.3154
ResNet101ERIG0.16170.65800.2419
SERIG0.25060.52390.3469
DeiT3ERIG0.11790.66370.1819
SERIG0.24820.49250.3372

Table 25. SERIG Ablation Localization: The signed variant wins IoU and PF1 on every backbone, while ERIG keeps the best PA.

ModelMethodMS (↓)LLC (↓)RIS (↓)
ViT-BaseERIG3.0e–049.0e–064.9460
SERIG6.1e–041.6e–0512.040
EVA02ERIG6.7e–051.5e–061.6010
SERIG4.5e–041.0e–0510.157
MobileNetV4ERIG1.0e–045.4e–062.1960
SERIG0.00135.0e–0521.166
EfficientNetERIG6.7e–052.1e–061.2710
SERIG7.2e–042.2e–0516.032
ResNet101ERIG8.6e–053.8e–061.0660
SERIG0.00421.9e–0441.584
DeiT3ERIG3.9e–051.0e–060.7560
SERIG0.00802.1e–0410.234

Table 26. SERIG Ablation Stability: ERIG is more stable in every cell, by up to two orders of magnitude.

ModelMethodLINV-C (↑)LINV-P (↑)
ViT-BaseERIG19.34319.201
SERIG12.02811.920
EVA02ERIG20.12820.115
SERIG12.63912.636
MobileNetV4ERIG18.13018.113
SERIG11.44011.331
EfficientNetERIG20.70520.712
SERIG11.97611.970
ResNet101ERIG16.80716.778
SERIG10.98710.932
DeiT3ERIG19.57819.584
SERIG11.68212.026

Table 27. SERIG Ablation Noise: ERIG attains the higher LINV in every cell, by 5.8 to 8.7 LINV over the signed variant.

changes in regions to which SERIG assigns little or no attribution. The stability cost has the same underlying cause. Since sign⁡(gi)\operatorname{sign}(g_i) is discontinuous at gi=0g_i = 0, small perturbations can flip the sign factor even when the corresponding change in the softmax magnitude is negligible.

The ablation also clarifies the role of sign at different levels of the attribution process. Sign at the level of the final attribution is clearly meaningful, and ERIG preserves this information. The signed δf\delta f increments indicate whether a feature’s importance is associated with increases or decreases in the class score along the integration path. For example, a feature that receives a small weight when δf\delta f is positive but a large weight when δf\delta f is negative can receive a negative attribution. This directional information is useful for downstream applications. In contrast, the sign of the instantaneous gradient is a property of an individual evaluation point. Because of gradient irregularity, this sign can change as the evaluation point moves along the integration path or across noise samples in SmoothGrad-style averaging. An aggregation based on the instantaneous gradient sign therefore reflects the behavior of the ensemble of evaluation points rather than an intrinsic property of the feature’s effect. The ablation suggests that this per-step gradient sign is not the primary determinant of explanation quality. Introducing it degrades fidelity, stability, and noise performance across all backbones, while improving object alignment. The broader comparison supports the same conclusion. Among the path-based methods evaluated, IDGI and ERIG, which discard the per-step gradient sign and retain sign information only through the increments, consistently achieve the strongest fidelity and stability. In contrast, methods that integrate the raw signed gradient, including IG, BIG, and GIG, perform worse on both criteria. Reintroducing the gradient sign into ERIG’s otherwise identical weighting reproduces this gap in a controlled setting. Because the signed variant requires no additional computational cost and uses the same weights, SERIG can be preferred for localization-oriented applications, while ERIG remains the default choice.

References

  1. [1]Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. AttnIrp: Attention-aware layer-wise relevance propagation for transformers. arXiv preprint arXiv:2402.05602, 2024. 2arxiv.org/abs/2402.05602
  2. [2]Amina Adadi and Mohammed Berrada. Peeking inside the black-box: a survey on explainable artificial intelligence (xai). IEEE access, 6:52138–52160, 2018. 1DOI
  3. [3]Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. Advances in neural information processing systems, 31, 2018. 1arxiv.org/abs/1810.03292
  4. [4]Chirag Agarwal, Nari Johnson, Martin Pawelczyk, Satyapriya Krishna, Eshika Saxena, Marinka Zitnik, and Himabindu Lakkaraju. Rethinking stability for attribution based explanations. arXiv preprint arXiv:2203.06877, 2022. 5, 6, 34arxiv.org/abs/2203.06877
  5. [5]David Alvarez-Melis and Tommi S Jaakkola. On the robustness of interpretability methods. arXiv preprint arXiv:1806.08049, 2018. 1, 5, 6, 34
  6. [6]Saltelli Andrea, Ratto Marco, Andres Terry, Campolongo Francesca, Cariboni Jessica, Gatelli Debora, Saisana Michaela, and T Stefano. Global sensitivity analysis: the primer. 1st John Wiley & Sons, The Atrium, Southern Gate, Chichester, England, 2008. 14
  7. [7]David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The shattered gradients problem: If resnets are the answer, then what is the question? In Proceedings of the International Conference on Machine Learning, pages 342–350, 2017. 1arxiv.org/abs/1702.08591
  8. [8]Naman Bansal, Chirag Agarwal, and Anh Nguyen. Sam: The sensitivity of attribution methods to hyperparameters. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition, pages 8673–8683, 2020. 1arxiv.org/abs/2003.08754
  9. [9]Umang Bhatt, Adrian Weller, and José MF Moura. Evaluating and aggregating feature-based model explanations. arXiv preprint arXiv:2005.00631, 2020. 31arxiv.org/abs/2005.00631
  10. [10]Timothy J. Boerner, Stephen Deems, Thomas R. Furlani, Shelley L. Knuth, and John Towns. Access: Advancing innovation: NSF’s advanced cyberinfrastructure coordination ecosystem: Services & support. In Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good, page 173–176, New York, NY, USA, 2023. Association for Computing Machinery. 9
  11. [11]Dan G Cacuci. Sensitivity and Uncertainty Analysis, Volume I: Theory. Chapman & Hall/CRC, 2003. 14
  12. [12]Prasad Chalassani, Jiefeng Chen, Amrita Roy Chowdhury, Xi Wu, and Shashi Jha. Concise explanations of neural networks using adversarial training. In International conference on machine learning, pages 1383–1391. PMLR, 2020. 31
  13. [13]Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 839–847. IEEE, 2018. 5DOI
  14. [14]Pratik Chaudhari, Anna Choromanska, Stefano Soatto, Yann LeCun, Carlo Baldassi, Christian Borgs, Jennifer Chayes, Levent Sagun, and Riccardo Zecchina. Entropy-sgd: Biasing gradient descent into wide valleys. Journal of Statistical Mechanics: Theory and Experiment, 2019(12):124018, 2019. 3arxiv.org/abs/1611.01838
  15. [15]Preeti S Chauhan and Nir Kshetri. The role of data and artificial intelligence in driving diversity, equity, and inclusion. Computer, 55(4):88–93, 2022. 1DOI
  16. [16]Guangyong Chen, Fengyuan Zhu, and Pheng Ann Heng. An efficient statistical method for image noise level estimation. In Proceedings of the IEEE international conference on computer vision, pages 477–485, 2015. 2, 7, 33DOI
  17. [17]Hugh Chen, Scott M Lundberg, and Su-In Lee. Explaining a series of models by propagating shapley values. Nature communications, 13(1):4512, 2022. 2DOI
  18. [18]Gabriela Csurka, Diane Larlus, Florent Perronnin, and France Meylan. What is a good evaluation measure for semantic segmentation?. In Bmvc, pages 10–5244. Bristol, 2013. 18DOI
  19. [19]Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. How important is a neuron. In International Conference on Learning Representations, 2018. 2DOI
  20. [20]Ann-Kathrin Dombrowski, Maximillian Alber, Christopher Anders, Marcel Ackermann, Klaus-Robert Müller, and Pan Kessel. Explanations can be manipulated and geometry is to blame. Advances in neural information processing systems, 32, 2019. 1, 23, 33arxiv.org/abs/1906.07983
  21. [21]John C Duchi, Peter L Bartlett, and Martin J Wainwright. Randomized smoothing for stochastic optimization. SIAM Journal on Optimization, 22(2):674–701, 2012. 23arxiv.org/abs/1103.4296
  22. [22]Joseph Enguehard. Sequential integrated gradients: a simple but effective method for explaining language models. In Findings of the Association for Computational Linguistics: ACL 2023, pages 7555–7565, 2023. 2arxiv.org/abs/2305.15853
  23. [23]Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving performance of deep learning models with axiomatic attribution priors and expected gradients. Nature machine intelligence, 3(7):620–631, 2021. 2, 29arxiv.org/abs/1906.10670
  24. [24]Christian Etmann, Sebastian Lunz, Peter Maass, and Carola-Bibiane Schönlieb. On the connection between adversarial robustness and saliency map interpretability. In International Conference on Machine Learning, pages 1823–1832. PMLR, 2019. 1arxiv.org/abs/1905.04172
  25. [25]Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 4DOI
  26. [26]Mahyar Fazlyab, Alexander Robey, Hamed Hassani, Manfred Morari, and George Pappas. Efficient and accurate estimation of lipschitz constants for deep neural networks. Advances in neural information processing systems, 32, 2019. 4arxiv.org/abs/1906.04893
  27. [27]Shanghua Gao, Zhong-Yu Li, Ming-Hsuan Yang, Ming-Ming Cheng, Junwei Han, and Philip Torr. Large-scale unsupervised semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 5DOI
  28. [28]Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI conference on artificial intelligence, pages 3681–3688, 2019. 1, 33arxiv.org/abs/1710.10547
  29. [29]Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1DOI
  30. [30]David Gunning and David Aha. Darpa’s explainable artificial intelligence (xai) program. AI magazine, 40(2):44–58, 2019. 1DOI
  31. [31]David Y. Hancock, Jeremy Fischer, John Michael Lowe, Winona Snapp-Childs, Marlon Pierce, Suresh Marru, J. Eric Coulter, Matthew Vaughn, Brian Beck, Nirav Merchant, Edwin Skidmore, and Gwen Jacobs. Jetstream2: Accelerating cloud computing via jetstream. In Practice and Experience in Advanced Research Computing 2021: Evolution Across All Dimensions, New York, NY, USA, 2021. Association for Computing Machinery. 9DOI
  32. [32]Boris Hanin and David Rolnick. Deep relu networks have surprisingly few activation patterns. Advances in neural information processing systems, 32, 2019. 1, 23arxiv.org/abs/1906.00904
  33. [33]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 4arxiv.org/abs/1512.03385
  34. [34]Pengcheng He, Jianfeng Gao, and Weizhu Chen. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing, 2021. 1arxiv.org/abs/2111.09543
  35. [35]Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Junyuan Xie, and Mu Li. Bag of tricks for image classification with convolutional neural networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 558–567, 2018. 4arxiv.org/abs/1812.01187
  36. [36]Anna Hedström, Leander Weber, Daniel Krakowczyk, Dilyara Bareeva, Franz Motzkus, Wojciech Samek, Sebastian Lapuschkin, and Marina Marina M.-C. Höhne. Quantus: An explainable ai toolkit for responsible evaluation of neural network explanations and beyond. Journal of Machine Learning Research, 24(34):1–11, 2023. 5arxiv.org/abs/2202.06861
  37. [37]Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. FunnyBirds: A synthetic vision dataset for a part-based analysis of explainable AI methods. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), Paris, France, October 2-6, 2023, pages 3981–3991. IEEE, 2023. 7, 25, 26arxiv.org/abs/2308.06248
  38. [38]Robin Hesse, Simone Schaub-Meyer, Janina Hesse, Bernt Schiele, and Stefan Roth. What is missing? explaining neurons activated by absent concepts. arXiv preprint arXiv:2603.09787, 2026. 34arxiv.org/abs/2603.09787
  39. [39]John Immerkaer. Fast noise variance estimation. Computer vision and image understanding, 64(2):300–302, 1996. 2, 7, 33DOI
  40. [40]Hyeongwon Jang, Changhun Kim, and Eunho Yang. TIM-ING: Temporality-Aware Integrated Gradients for Time Series Explanation. In International Conference on Machine Learning (ICML), 2025. 2
  41. [41]Edwin T Jaynes. Information theory and statistical mechanics. Physical review, 106(4):620, 1957. 3
  42. [42]Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4948–4957, 2019. 2, 4, 5arxiv.org/abs/1906.02825
  43. [43]Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided integrated gradients: An adaptive path method for removing noise. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5050–5058, 2021. 2, 4, 5, 6, 7, 29, 34arxiv.org/abs/2106.09788
  44. [44]Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un)reliability of saliency methods. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267–280. Springer, 2019. 1arxiv.org/abs/1711.00867
  45. [45]Rohith Krishna, Jue Wang, Woody Ahern, Pascal Sturmfels, Preetham Venkatesh, Indrek Kalvet, Gyu Rie Lee, Felix S Morey-Burrows, Ivan Anishchenko, Ian R Humphreys, et al. Generalized biomolecular modeling and design with rosettafold all-atom. Science, 384(6693):eadl2528, 2024. 1
  46. [46]Matieyendou Lamboni, Bertrand Iooss, Anne-Laure Popelin, and Fabrice Gamboa. Derivative-based global sensitivity measures: General links with Sobol’ indices and numerical tests. Mathematics and Computers in Simulation, 87:45–54, 2013. 14arxiv.org/abs/1202.0943
  47. [47]Shuyang Liu, Zixuan Chen, Ge Shi, Ji Wang, Changjie Fan, Yu Xiong, Runze Wu Yujing Hu, Ze Ji, and Yang Gao. A new baseline assumption of integrated gradients based on shapley value. arXiv preprint arXiv:2310.04821, 2023. 2, 29arxiv.org/abs/2310.04821
  48. [48]Faridoun Mehri, Mahdieh Soleymani Baghshah, and Mohammad Taher Pilehvar. Libragrad: Balancing gradient flow for universally better vision transformer attribution. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 67–78. IEEE, 2025. 2arxiv.org/abs/2411.16760
  49. [49]Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Jonathan Uesato, and Pascal Frossard. Robustness via curvature regularization, and vice versa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9078–9086, 2019. 1arxiv.org/abs/1811.09716
  50. [50]Rukmangadh Sai Myana and Sumit Kumar Jha. Explaining protein folding networks using integrated gradients and attention mechanisms. In Computational Advances in Bio and Medical Sciences, pages 195–211, Cham, 2026. Springer Nature Switzerland. 1DOI
  51. [51]Pravin Nair. Softmax is 1/2-Lipschitz: A tight bound across all ℓp norms. arXiv preprint arXiv:2510.23012, 2025. 21
  52. [52]Random gradient-free minimization of convex functions. Foundations of Computational Mathematics, 17(2):527–566, 2017. 23
  53. [53]Explaining deep neural network models with adversarial gradient integration. In Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), 2021. 4, 5
  54. [54]Fairness in mobile phone–based mental health assessment algorithms: Exploratory study. JMIR formative research, 6(6):e43466, 2022. 1
  55. [55]Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498–3505. IEEE, 2012. 7, 25
  56. [56]Computational optimal transport. Foundations and Trends in Machine Learning, 51 (1):1–44, 2019. 3arxiv.org/abs/1803.00567
  57. [57]MobileNetv4-universal models for the mobile ecosystem. arXiv preprint arXiv:2404.10518, 2024. 4
  58. [58]Irof: a low resource evaluation metric for explanation methods. arXiv preprint arXiv:2003.08747, 2020. 5
  59. [59]ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211–252, 2015. 5arxiv.org/abs/1409.0575
  60. [60]Agi-accelerated nazca survey nearly doubles the number of known figurative geoglyphs and sheds light on their purpose. Proceedings of the National Academy of Sciences, 121 (40):e2407652121, 2024. 1
  61. [61]How does batch normalization help optimization? In Advances in Neural Information Processing Systems, 2018. 1DOI
  62. [62]On the explainable properties of 1-lipschitz neural networks: An optimal transport perspective. Advances in Neural Information Processing Systems, 36:54645–54682, 2023. 4
  63. [63]Deep learning models for water stage predictions in south florida. Journal of Water Resources Planning and Management, 151(10):05025013, 2025. 1
  64. [64]Deep learning and foundation models for weather prediction: A survey. In Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, IJCAI-26, pages 8023–8031. International Joint Conferences on Artificial Intelligence Organization, 2026. Survey Track. 1
  65. [65]Tangentially aligned integrated gradients for user-friendly explanations. arXiv preprint arXiv:2503.08240, 2025. 2, 29
  66. [66]Riemann sum optimization for accurate integrated gradients computation. In Interpretable AI: Past, Present and Future, 2024. 2, 29
  67. [67]When explanations lie: Why many modified bp attributions fail. In International conference on machine learning, pages 9046–9057. PMLR, 2020. 2arxiv.org/abs/1912.09818
  68. [68]Smoothgrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. 1, 3, 5, 7, 23, 24arxiv.org/abs/1706.03825
  69. [69]Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates. Mathematics and computers in simulation, 55(1-3):271–280, 2001. 14
  70. [70]Derivative based global sensitivity measures and their link with global sensitivity indices. Mathematics and Computers in Simulation, 79(10):3009–3017, 2009. 3, 14
  71. [71]Efficiently training low-curvature neural networks. In Advances in Neural Information Processing Systems, 2022. 1arxiv.org/abs/2206.07144
  72. [72]Axiomatic attribution for deep networks. In International conference on machine learning, pages 3319–3328. PMLR, 2017. 1, 2, 4, 5, 16
  73. [73]Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1arxiv.org/abs/1312.6199
  74. [74]Compensated integrated gradients to reliably interpret eeg classification. arXiv preprint arXiv:1811.08633, 2018. 2, 29
  75. [75]Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105–6114. PMLR, 2019. 4
  76. [76]Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 4arxiv.org/abs/2204.07118
  77. [77]Robustness may be at odds with accuracy. In International Conference on Learning Representations, 2019. 1
  78. [78]Aladin Virmaux and Kevin Scaman. Lipschitz regularity of deep neural networks: analysis and efficient estimation. Advances in neural information processing systems, 31, 2018. 4arxiv.org/abs/1805.10965
  79. [79]Chase Walker, Sumit Jha, Kenny Chen, and Rickard Ewetz. Integrated decision gradients: Compute your attributions where the model makes its decision. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5289–5297, 2024. 2arxiv.org/abs/2305.20052
  80. [80]Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984–108011, 2024. 1arxiv.org/abs/2405.14458
  81. [81]Zhou Wang, Eero P. Simoncelli, and Alan C Bovik. Multi-scale structural similarity for image quality assessment. In The thirty-seventh asilomar conference on signals, systems & computers, 2003, pages 1398–1402. Ieee, 2003. 5
  82. [82]Zifu Wang, Maxim Berman, Amal Rannen Triki, Philip Torr, Devis Tuia, Tinne Tuytelaars, Luc V Gool, Jiaqian Yu, and Matthew Blaschko. Revisiting evaluation metrics for semantic segmentation: Optimization and evaluation of fine-grained intersection over union. Advances in Neural Information Processing Systems, 36:60144–60225, 2023. 18arxiv.org/abs/2310.19252
  83. [83]Zichong Wang, Jocelyn Dzuong, Xiaoyong Yuan, Zhong Chen, Yanzhao Wu, Yin Yao, and Wenbin Zhang. Individual fairness with group awareness under uncertainty. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 89–106. Springer, 2024. 1DOI
  84. [84]Wenqi Wei, Ling Liu, Margaret Loper, Ka-Ho Chow, Emre Gursory, Stacey Truex, and Yanzhao Wu. Cross-layer strategic ensemble defense against adversarial examples. In 2020 International Conference on Computing, Networking and Communications (ICNC), pages 456–460. IEEE, 2020. 1arxiv.org/abs/1910.01742
  85. [85]Ross Wightman. Pytorch image models. https://github.com/huggingface/pytorch-image-models, 2019. 5
  86. [86]Sangyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16133–16142, 2023. 1arxiv.org/abs/2301.00808
  87. [87]Adam Wróbel, Mikołaj Janusz, Bartosz Zieliński, and Dawid Rymarczyk. Omenn: One matrix to explain neural networks. arXiv preprint arXiv:2412.02399, 2024. 2arxiv.org/abs/2412.02399
  88. [88]Adam Wróbel, Siddhartha Gairola, Jacek Tabor, Bernt Schiele, Bartosz Zieliński, and Dawid Rymarczyk. Dave: Distribution-aware attribution via git gradient decomposition. arXiv preprint arXiv:2602.06613, 2026. 2arxiv.org/abs/2602.06613
  89. [89]Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision, 2020. 4arxiv.org/abs/2006.03677
  90. [90]Yanzhao Wu, Ling Liu, Zhongwei Xie, Ka-Ho Chow, and Wenqi Wei. Boosting ensemble accuracy by revisiting ensemble diversity metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16469–16477, 2021. 1DOI
  91. [91]Yanzhao Wu, Ka-Ho Chow, Wenqi Wei, and Ling Liu. Exploring model learning heterogeneity for boosting ensemble robustness. In 2023 IEEE International Conference on Data Mining (ICDM), pages 648–657. IEEE, 2023. 1arxiv.org/abs/2310.02237
  92. [92]Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9680–9689, 2020. 2, 5, 29arxiv.org/abs/2004.03383
  93. [93]Ruo Yang, Binghui Wang, and Mustafa Bilgic. Idgi: A framework to eliminate explanation noise from integrated gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23725–23734, 2023. 1, 2, 5, 6, 7, 34arxiv.org/abs/2303.14242
  94. [94]Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Suggala, David I Inouye, and Pradeep K Ravikumar. On the (in) fidelity and sensitivity of explanations. Advances in neural information processing systems, 32, 2019. 1, 5, 6, 34arxiv.org/abs/1901.09392
  95. [95]Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 1arxiv.org/abs/1906.08988
  96. [96]Eslam Zaher, Maciej Trzaskowski, Quan Nguyen, and Fred Roosta. Manifold integrated gradients: Riemannian geometry for feature attribution. In Proceedings of the 41st International Conference on Machine Learning, pages 58090–58104, 2022. 2
  97. [97]Yue Zhuo and Zhiqiang Ge. Ig2: Integrated gradient on iterative gradient path for feature attribution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(11):7173–7190, 2024. 2, 29DOI
  98. [98]Zhike Zi. Sensitivity analysis approaches applied to systems biology models. IET Systems Biology, 5(6):336–346, 2011. 14DOI

Paper details

vlnc:W_ry9m5cps
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents