WorksComputer Science
Spiking Neural Networks Are Harder to Steal: A Benchmark Study of Model Stealing Attacks
Abstract
Machine learning models deployed behind query interfaces expose more than their predictions, because an adversary can use those predictions to rebuild the model itself. Model stealing attacks (MSAs) exploit this exposure by training a surrogate model that replicates the decision behavior of a target through black-box queries alone, without access to its parameters or training data. These attacks are well characterized for conventional artificial neural networks (ANNs), yet spiking neural networks (SNNs) process information in a fundamentally different way. SNNs communicate through sparse, discrete spikes, and their outputs depend on membrane potentials that evolve over time rather than on a static mapping from input to output. This temporal mode of computation suggests that the decision behavior of an SNN may be harder to reconstruct from its responses. This work tests that possibility by benchmarking four representative MSAs, namely the Papernot, Tramèr, T-RND, and color perturbation attacks, against SNNs and their ANN counterparts on MNIST, Fashion-MNIST, and CIFAR-10. Across every attack and dataset, SNNs prove consistently more difficult to steal. Surrogates of SNNs achieve 10% to 20% lower agreement with the target model and 15% to 25% lower adversarial transferability than surrogates of ANNs. We attribute this resistance to the way SNNs compute, as their spike-based and time-dependent dynamics yield decision boundaries that ANN-based surrogates struggle to approximate, even under large query budgets. These findings extend the known resilience of SNNs beyond attacks on training data to the replication of model functionality, strengthening the case for neuromorphic architectures in security-sensitive applications.
Introduction
Machine learning (ML) models are commonly used in sensitive applications, raising questions about how their decision behavior can be protected against adversarial exploitation [27]. Model stealing attacks (MSAs) represent a particularly consequential threat in ML models that are accessible to users. An adversary with only query access to a model can systematically reconstruct its decision behavior and extract the model’s underlying intellectual property embedded in its parameters and architecture [4, 18]. Beyond the loss of proprietary value and the cost of model development, extracted models can serve as high-fidelity proxies for further attacks, including adversarial example transfer and membership inference, amplifying the downstream security risk to the original system [24]. As ML is increasingly adopted in sensitive domains such as health-care [22], finance [2], and edge computing [30], it is important to understand which architectural choices naturally resist model stealing through black-box query access.
MSA, formally introduced by Tramèr et al. [27] and extended in subsequent work [7, 17], operates by iteratively querying a target or victim model which refers to the original deployed model accessible only through input-output queries. The attacker uses these predictions to train a substitute or surrogate model that approximates the target’s decision function. The attacker requires no knowledge of the model’s internal parameters, training data, or architecture. Given sufficient queries and an appropriate strategy for generating informative inputs, surrogate models trained in this manner can achieve high agreement with the victim and effectively inherit its vulnerabilities. Prior works have demonstrated that such attacks are feasible across a range of model types and datasets, and that stronger query strategies, particularly those that focus on decision boundary regions, substantially improve extraction fidelity [3, 18]. However, existing studies are largely limited to conventional artificial neural networks (ANNs), with little investigation into whether alternative architectures with fundamentally different computational properties exhibit greater resistance to extraction.
Spiking Neural Networks (SNNs) represent one such alternative. Inspired by biological neural systems, SNNs have attracted attention for their energy efficiency and ability to process temporal information, and are often implemented on neuromorphic hardware designed for event-driven computation. Unlike ANNs, which propagate dense, real-valued activations, SNNs encode information through discrete spike trains whose timing and sparsity carry the representational signal. Computation in SNNs is governed by the membrane potential dynamics of spiking neurons, where each neuron’s response depends on the accumulated history of its inputs over time rather than a static input-output mapping, as seen in models such as leaky integrate-and-fire (LIF) neurons. These architectural differences suggest that the decision boundaries learned by SNNs may be fundamentally harder to approximate through black-box query access, motivating an examination of their resistance to model stealing.
Recent studies have begun to examine the privacy properties of SNNs from multiple perspectives. Research on Membership Inference Attacks (MIA) [24] indicate that SNNs leak less information about their training data compared to ANNs [12], while model inversion studies by Poursiami et al. demonstrate greater resistance to reconstructing input data from outputs in both white-box and black-box settings [20, 21]. These findings suggest that the sparse, event-driven, and temporally dynamic nature of SNN computation may inherently limit information leakage. However, prior work on privacy attacks, such as MIA, has primarily focused on data-centric threats, evaluating whether individual samples can be inferred or reconstructed. Whether these advantages extend to MSAs has not been evaluated. Unlike MIAs, which aim to determine whether specific training samples were used during training, MSAs aim to replicate the functional behavior of a model through query access without recovering the underlying training data.
This work investigates whether the privacy advantages observed in SNNs extend to resistance against MSAs. We present a benchmark of model extraction attacks on SNNs. We evaluate four MSA strategies, namely gradient-based methods such as the Papernot attack [18], boundary-focused extraction by Tramèr et al. [27], as well as query-based strategies [8, 17] including TRND and input-space color perturbations [3, 8], across the datasets: MNIST [10], FMNIST [28], and CIFAR10 [9]. Extraction performance is evaluated along three dimensions: (i) functional fidelity, (ii) adversarial transferability, and (iii) query efficiency. Functional fidelity is measured through test and random uniform agreement, while adversarial transferability is assessed using both non-targeted and targeted metrics [19, 26]. Query efficiency is evaluated based on the number of queries required to achieve a given level of agreement. Across all settings, SNNs consistently exhibit lower extractability than ANNs, with surrogate agreement reduced by 10–20% and transferability by 15–25%, even under attacks with large query budgets. This behavior can be attributed to the intrinsic properties of SNNs [6, 25], where sparse spike-based communication and temporal membrane dynamics produce input-output relationships that are more difficult to approximate through static query-response pairs. This gap persists even when the surrogate model is also an SNN, which indicates that the resistance arises from the target model’s temporal computation rather than differences between the architectures of the surrogate and target models.
The main contributions of this paper are as follows:
SNNs exhibit stronger resistance to model stealing. Across four attacks, namely Papernot, Tramèr, T-RND, and COLOR, and three datasets: MNIST, FMNIST, and CIFAR-10, SNN targets show 10–20% lower surrogate agreement and 15–25% lower adversarial transferability than ANNs. This suggests that their temporal and sparse representations are more difficult to approximate through black-box queries.
Resistance arises from intrinsic dynamics. Low transferability persists even when the surrogate is an SNN (approximately 0.21–0.23). This shows that the effect is driven by the target model’s temporal and sparse computation rather than architectural mismatch.
Extraction is less query-efficient for SNNs. ANN surrogates reach high accuracy within a few thousand queries, whereas SNN surrogates improve gradually even at large budgets. This reflects slower convergence during extraction.
Temporal and sparse computation limits extractability. Spike-based communication and history-dependent neuron dynamics reduce the usable signal in query responses. As a result, the learned decision function is harder to reconstruct.
Background
Privacy attacks on ML models have been studied across several threat models. Model inversion attacks recover representative inputs or sensitive attributes from model predictions [5], while MIAs determine whether specific samples were used during training [24, 29]. These attacks have been demonstrated against a range of ANN architectures and establish that black-box query access alone is sufficient to expose information about a model’s training data. In the context of SNNs, prior work has shown that these architectures exhibit reduced vulnerability under both threat models, with lower MIA success rates [12, 13, 14, 15] and greater resistance to input reconstruction from outputs [20, 21]. These findings are generally attributed to the sparse, event-driven nature of spike-based computation and the temporal dynamics of LIF neurons. However, this body of work has focused exclusively on data-centric threats, where the objective is to recover information about training samples. Whether the same architectural properties limit an adversary’s ability to replicate the model’s decision function through query access has not been examined. This distinction motivates a closer examination of SNN computation and model stealing mechanisms to understand how architectural properties influence extractability under black-box query access.
Spiking Neural Networks (SNN)
SNNs encode information through discrete spike events over time rather than continuous activations. Neurons integrate incoming signals into a membrane potential and emit a spike when this potential exceeds a threshold, resulting in sparse and temporally structured representations. This event-driven paradigm has been widely studied in computational neuroscience and neuromorphic computing [6, 11].
A commonly used neuron model is the LIF formulation, where the membrane potential evolves as:
where denotes the membrane potential, is the input current, and is the membrane time constant. When exceeds a threshold , the neuron emits a spike and the potential is reset [1, 6].
In discrete-time implementations commonly used in SNNs, the dynamics are approximated as:
where denotes the spike output and represents the leakage factor. This formulation captures the temporal accumulation of inputs and the dependence of neuron responses on prior activity [16].
Unlike conventional architectures, computation in SNNs depends on both current inputs and the evolution of membrane potentials over time. Combined with sparse spike-based communication, this leads to input-output relationships that are mediated by internal dynamics and are not fully observable through individual query responses. As a result, approximating SNN behavior from black-box access may require capturing temporal dependencies that are absent in standard feedforward models [23].
Model Stealing Attacks (MSA)
MSAs, also referred to as model extraction, aim to replicate the functionality of a target model through black-box query access. An adversary interacts with the model by submitting inputs and observing outputs, and uses this information to train a surrogate model that approximates the target’s decision function. This process enables the recovery of model behavior without access to internal parameters or training data.
In addition to replicating functionality, extracted models can serve as proxies for further attacks, including adversarial example transfer and membership inference [7]. In this work, we use MSAs to evaluate whether SNNs are more difficult to replicate than ANN counterparts under comparable query budgets and attack strategies. To this end, we consider 4 attacks: the Papernot attack [18], the Tramèr attack [27], and the T-RND and COLOR attack variants evaluated in [8].
Papernot Attack
The Papernot attack [18] constructs a substitute or surrogate model by iteratively querying a target classifier and augmenting the training data using gradient-based perturbations. Let denote the target model, which is accessible only through black-box queries, and denote the substitute model.
The attack begins with a small seed input set . At each iteration , the current input set is labeled by querying the target model to form a labeled dataset
The substitute model is then trained on to approximate the behavior of .
New samples are generated by perturbing inputs in using the gradient of the substitute model:
where controls the perturbation magnitude. Since the attacker does not have access to the gradients of the target model, gradients are computed with respect to the substitute model .
The input set is then updated as
This process of querying, training, and augmentation is repeated until the query budget is exhausted (Figure 1). This iterative procedure enables the substitute model to progressively approximate the decision boundaries of the target model, improving both predictive agreement and adversarial transferability.
Figure 1. Training of the substitute model . The attacker first collects an initial input set and selects a substitute architecture . At each iteration , the attacker queries the oracle model on to obtain labeled data , which is used to train the substitute model. Synthetic samples are generated via gradient-based augmentation, , and added to the input set to form . This process is repeated iteratively.
Tramèr Attack
The Tramèr attack [27] improves query efficiency by generating samples near the decision boundaries of the target model rather than relying on uniform augmentation. Starting from an initial labeled dataset
a substitute model is trained to approximate the target model . At each iteration , synthetic samples are generated by applying an iterative fast gradient sign method to inputs in :
where controls the perturbation magnitude and denotes the number of steps. This procedure moves samples toward the decision boundary of the substitute model.
The labeled dataset is then updated as
The substitute model is retrained on the expanded dataset, and the process is repeated until the query budget is exhausted (Figure 2). By concentrating queries near decision boundaries, this approach enables faster approximation of the target decision function and improves extraction efficiency.
Figure 2. Overview of the Tramèr model extraction procedure. The attacker first constructs an initial labeled dataset by querying the oracle model . A substitute model is trained on the current dataset . Synthetic samples are then generated near the decision boundary using an iterative fast gradient sign method, , and labeled via oracle queries. The dataset is updated as , and the process is repeated for successive iterations until the query budget is exhausted.
TRND Attack
The TRND attack extends boundary-focused query generation by incorporating targeted perturbations with random class selection, enabling broader exploration of the decision space. Starting from an initial labeled dataset
a substitute model is trained to approximate the target model . At each iteration , synthetic samples are generated by applying a targeted iterative fast gradient sign method. For an input , a target class is randomly selected such that , and the input is perturbed as
where controls the perturbation magnitude and denotes the number of steps. This procedure drives samples toward regions associated with different target classes.
The labeled dataset is then updated as
The substitute model is retrained on the expanded dataset until the query budget is exhausted.
By combining targeted perturbations with random class selection, this approach improves coverage of the decision space compared to untargeted methods, leading to more effective model extraction.
Color Perturbation Attack
The Color attack generates synthetic queries through input-space transformations rather than gradient-based perturbations. Starting from an initial labeled dataset
a substitute model is trained to approximate the target model . At each iteration , new samples are produced by applying random color or intensity transformations to inputs in :
where denotes a transformation function that perturbs input statistics, such as brightness, contrast, or channel-wise scaling, and controls the perturbation magnitude.
The labeled dataset is then updated as
The substitute model is retrained on the expanded dataset until the query budget is exhausted (Figure 3).
Figure 3. Color perturbation examples on an FMNIST sample. The transformations preserve semantic structure while modifying input statistics for query generation.
Unlike gradient-based methods, this approach does not rely on decision boundary information and instead emphasizes query diversity through input-space variation, providing a lightweight baseline for model extraction.
Experimental Setup
Experiments are conducted to evaluate MSAs across both ANN and SNN targets under a unified experimental setting. The objective is to compare extractability across architectures while keeping datasets, training procedures, and attack configurations fixed. Evaluations are performed on MNIST, FMNIST, and CIFAR10, covering increasing levels of dataset complexity.
Model Architecture & Datasets
Datasets: Three image classification benchmarks are considered: MNIST [10], FMNIST [28], and CIFAR10 [9]. These datasets span increasing levels of complexity, from grayscale digit recognition to natural image classification, and are used to evaluate extraction behavior under different data conditions.
ANN Architectures: For ANN targets and substitutes, convolutional neural networks (CNNs) are employed, consisting of two convolutional layers followed by fully connected layers. Each convolutional layer is followed by a ReLU activation and max pooling. The architectures are kept consistent across datasets, with input channels adjusted for grayscale or RGB inputs. Models are trained using cross-entropy loss with the Adam optimizer, a learning rate of , and a batch size of 128. Standard data normalization is applied during training.
| Layer | MNIST / FMNIST | CIFAR10 |
| Input Channels | 1 | 3 |
| Conv Layer 1 + ReLU | 32, | 32, |
| MaxPool | ||
| Conv Layer 2 + ReLU | 64, | 64, |
| MaxPool | ||
| Flattened Features | ||
| Fully Connected + ReLU | 1000 | 1000 |
| Output Layer | 10 | 10 |
Table 1. ANN Architectures
SNN Architectures: SNN models are constructed using a similar convolutional backbone, with standard activations replaced by spiking neurons governed by leaky integrate-and-fire (LIF) dynamics. Convolutional and fully connected layers are followed by LIF neurons. Inputs are encoded into spike trains using rate-based encoding and processed over discrete time steps. Neuron dynamics include membrane potential accumulation with leakage factor , threshold-based spike generation, and reset after firing. Outputs are computed from the final membrane potential at the last time step.
| Layer | MNIST / FMNIST | CIFAR10 |
| Input Channels | 1 | 3 |
| Conv Layer 1 + LIF | 32, | 32, |
| MaxPool | ||
| Conv Layer 2 + LIF | 64, | 64, |
| MaxPool | ||
| Flattened Features | ||
| Fully Connected + LIF | 1000 | 1000 |
| Output + LIF | 10 | 10 |
| Time Steps | 25 | 25 |
| Leakage | 0.95 | 0.95 |
Table 2. SNN Architectures
For SNNs, neuron dynamics are governed by membrane potential accumulation with leakage and threshold-based spike generation. The models are trained using surrogate gradient methods with cross-entropy loss applied to spike count outputs, and optimized using Adam with a learning rate of .
Target Model Training: All target models are trained from scratch using backpropagation. Optimization is performed using mini-batch training for 50 epochs. For SNNs, training is carried out over multiple discrete time steps, and gradients are approximated using surrogate gradient methods.
Data preprocessing follows standard practices for each dataset. CIFAR10 training includes random cropping and horizontal flipping for data augmentation. MNIST and FMNIST inputs are resized to , converted to grayscale where applicable, and normalized.
Attack Setup
All attacks are performed in a black-box setting, where access to the target model is limited to input-output queries. Each attack begins with a small balanced seed set containing 10 natural samples per class, resulting in 100 initial queries for MNIST, FMNIST, and CIFAR10. These queried samples are used to form the initial labeled set for training the substitute model.
After the initial seed stage, the attack proceeds iteratively. In each duplication round, new synthetic queries are generated from the current labeled data, submitted to the target model, and added to the substitute training set. The substitute model is then retrained on the expanded dataset. This process continues either until 10 duplication rounds have been completed or until the maximum query budget of 102,400 is reached. After seed acquisition and after each subsequent duplication round, the substitute model is trained for 10 epochs.
The four attacks differ in how synthetic queries are generated. The Papernot attack uses Jacobian-based augmentation, where new inputs are generated by perturbing samples along the gradient of the substitute model with respect to the input, approximating directions of high sensitivity in the model’s predictions. The Tramèr attack generates queries near decision boundaries using iterative FGSM with perturbation budget and 5 update steps. The T-RND attack extends this approach by applying targeted random iterative FGSM with and 11 steps, producing queries that are pushed toward randomly selected target classes. The Color attack does not rely on gradients. Instead, inputs are perturbed directly in the input space using and an expansion factor of 4, producing four transformed variants per sample.
To maintain comparability, all attacks are evaluated under the same seed construction, substitute training schedule, and overall query budget for both ANN and SNN targets. In addition to extraction fidelity, transfer-based evaluation is performed on 1,000 held-out test samples, while global agreement is measured on 4,000 uniformly sampled random inputs. For transferability evaluation, adversarial examples are generated with perturbation budget and 11 iterative steps in both targeted and non-targeted settings.
Evaluation Metrics
The effectiveness of a MSA is not determined solely by predictive accuracy, but by how closely the surrogate or substitute model reproduces the functional behavior and vulnerabilities of the target model. To capture this, 4 evaluation metrics are used, where each reflects a distinct aspect of surrogate fidelity and behavioral alignment.
Test Agreement (Fidelity on In-Distribution Data). The primary objective of model stealing is to approximate the decision function of the target model. Test agreement measures the extent to which the surrogate produces the same predictions as the target on unseen in-distribution test samples, where in-distribution refers to data drawn from the same distribution as the training data (i.e., the standard test split of the dataset). For SNNs, predictions are obtained from the final membrane potential, while for ANNs they are obtained directly from the output logits.
Formally, for a test set , agreement is defined as
High test agreement indicates that the surrogate model has successfully captured the behavior of the target model on standard test inputs.
Random Uniform Agreement (Global Decision Alignment). While test agreement evaluates fidelity on the data distribution, it does not reveal whether the surrogate has captured the broader structure of the target model’s decision function. Random uniform agreement addresses this by measuring agreement on inputs sampled uniformly from the normalized input range.
Agreement is computed as
where each is sampled uniformly from the normalized input domain. High agreement in this setting suggests that the surrogate has learned a more global approximation of the target function.
Non-Targeted Transferability (Shared Vulnerability Structure). A successful MSA should not only reproduce predictions, but also capture the vulnerabilities of the target model. Non-targeted transferability measures whether adversarial examples crafted on the surrogate induce misclassification in the target model.
Formally, given adversarial examples generated on the surrogate,
In the implementation, adversarial examples are generated on the surrogate using iterative gradient-based methods. Higher values indicate stronger alignment in vulnerability structure between the surrogate and target models.
Targeted Transferability (Fine-Grained Behavioral Control). Targeted transferability evaluates whether adversarial examples crafted on the surrogate can drive the target model toward specific target labels. For each input, a target class different from the original prediction is selected.
The metric is defined as
Higher targeted transferability indicates that the surrogate captures fine-grained directional structure in the target model’s decision boundary.
Evaluation
MSA is evaluated along three dimensions: (i) fidelity, (ii) adversarial transferability, and (iii) query efficiency. Fidelity measures how closely the surrogate matches the target, while transferability evaluates whether the surrogate captures the target’s decision boundary and vulnerabilities. Query efficiency reflects the cost of extraction. The results reported in this section are obtained using ANN-based surrogate models for both ANN and SNN targets, ensuring a consistent basis for comparison.
Fidelity of Model Extraction
Test Agreement
Figure 4a shows the agreement between surrogate and target models on test data. High agreement is consistently observed for ANN targets across all datasets and attack methods, confirming that MSA can effectively replicate their decision behavior. For instance, on MNIST, TRAMER and PAPERNOT achieve agreement close to or above 0.9 for ANN, while SNN are consistently lower by approximately 10–20%. A similar gap is observed across FMNIST and CIFAR10, indicating a stable difference in extractability.
Figure 4. Fidelity of model extraction under MSA. Test agreement evaluates in-distribution fidelity, while random uniform agreement assesses global decision alignment.
Across all datasets, SNN consistently exhibit reduced agreement compared to ANNs. This gap persists even for stronger attacks such as TRAMER and PAPERNOT, suggesting that SNNs are inherently more difficult to extract. As dataset complexity increases from MNIST to CIFAR10, agreement decreases for both architectures, but the relative gap between ANN and SNN targets remains stable.
Random Uniform Agreement
To assess global alignment, Figure 4b reports agreement on randomly sampled inputs. Agreement drops for all models compared to the test setting, indicating that surrogate learning is largely confined to the data distribution. This drop is significantly more pronounced for SNNs. For example, while ANN targets retain moderate agreement under random inputs (typically around 0.5–0.7 for stronger attacks), SNN are lower by approximately 15–25% across all methods. This consistent gap further highlights the difficulty of capturing the global decision structure of SNN models.
The same ranking of attack methods is preserved, with TRAMER and PAPERNOT outperforming others, but even these methods struggle to achieve strong global alignment for SNNs. Overall, these results indicate that ANN models allow more globally consistent extraction, whereas SNN extraction remains largely localized.
Adversarial Transferability
Non-Targeted Transfer
Figure 5a shows non-targeted adversarial transferability across datasets and attack methods. For ANNs, attacks that actively explore decision boundaries, such as the Tramèr and Papernot attack, achieve high transferability, reaching approximately 0.85–0.95 on MNIST and around 0.75–0.85 on FMNIST. In contrast, SNNs are consistently lower by roughly 15–25%, typically remaining in the range of 0.60–0.75 on MNIST. and 0.50–0.65 on FMNIST. A similar gap persists on CIFAR10, where ANN transferability is around 0.60–0.70, while SNN targets drop closer to 0.40–0.55.
Figure 5. Adversarial transferability under MSA. Non-targeted transfer evaluates general vulnerability alignment, while targeted transfer assesses the ability to induce specific misclassifications.
In contrast, methods that rely on input-space transformations or less structured query generation, such as the Color perturbation attack and the T-RND attack, show substantially reduced transferability overall, often below 0.50 for ANN targets and dropping further to 0.30–0.45 for SNNs. Despite this, the relative gap between ANN and SNN targets remains consistent across all methods. These results indicate that surrogate models are able to capture decision boundaries of ANN targets more effectively, while SNN decision boundaries remain harder to approximate and exploit.
Targeted Transfer
Targeted Transfer. Figure 5b presents targeted transferability, a stricter evaluation where adversarial examples must induce specific incorrect predictions.
Overall transferability is lower compared to the non-targeted setting. For ANNs, the Tramèr attack and the Papernot attack achieve around 0.45–0.60 on MNIST and approximately 0.35–0.50 on FMNIST. In contrast, SNN targets are reduced by approximately 20–30%, typically falling within 0.20–0.40 across datasets. On CIFAR10, targeted transfer further decreases, with ANN targets around 0.25–0.40 and SNN targets often below 0.25. The larger gap in this setting suggests that SNN models are not only harder to approximate, but also more resistant to precise adversarial manipulation. Even when
surrogate fidelity is reasonable, it does not translate into strong targeted transfer for SNNs.
Query Efficiency
Figure 6 illustrates surrogate performance as a function of query budget for the Papernot and Tramèr attacks on FMNIST.
Figure 6. Query efficiency on FMNIST for Papernot and Tramer attacks. ANN targets achieve high accuracy with fewer queries, while SNN targets require larger budgets and still exhibit lower performance.
For ANN targets, both methods rapidly achieve high accuracy, exceeding 90% with relatively few queries (approximately 2k–6k), after which performance saturates. In contrast, SNN targets require substantially more queries and still remain lower by approximately 10–20% across all budgets.
As the number of queries increases up to 100k, ANN performance plateaus early, whereas SNN performance improves more gradually, indicating slower approximation of the target decision behavior. This suggests that increasing query budgets yields diminishing returns for SNN extraction.
Overall, these results indicate that SNN models are not only harder to extract in terms of final performance, but also less query-efficient, requiring more queries to achieve comparable levels of agreement.
Effect of Target and Substitute Architectures
To analyze the impact of architectural mismatch between surrogate and target models, we evaluate performance across different combinations of ANN and SNN models on MNIST dataset.
Figure 7 compares performance across different combinations of substitute and target architectures. For test agreement, both ANN and SNN substitutes achieve high fidelity when approximating ANN targets (around 0.90–0.91), while agreement is slightly lower for SNN targets (approximately 0.88–0.90). However, a larger difference emerges in transferability. For ANN targets, non-targeted transferability reaches around 0.40 when using an ANN substitute and remains comparable (around 0.31) with an SNN substitute. In contrast, for SNN targets, transferability drops to approximately 0.23–0.21 regardless of the substitute type. These results indicate that while surrogate architecture has limited impact on agreement, the target architecture plays a dominant role in determining adversarial transferability. In particular, SNN targets exhibit consistently lower transferability, even when surrogate fidelity is comparable. This suggests that the reduced vulnerability of SNN models is not due to the choice of substitute architecture, but rather due to intrinsic differences in their decision behavior.
Figure 7. Effect of target and substitute architectures on test agreement and non-targeted transferability on MNIST.
Conclusion
MSAs pose a practical and well-demonstrated threat to deployed machine learning systems, yet their effectiveness has been studied almost exclusively against conventional ANNs. This work examined whether SNNs, whose prior privacy advantages under membership inference and model inversion attacks suggest a broader resilience, also resist functional replication through black-box query access. Across all attack strategies, query budgets, and datasets evaluated, SNNs consistently exhibit lower surrogate agreement and reduced adversarial transferability compared to ANNs. This behavior is rooted in the underlying computation of SNNs. Sparse spike-based communication and temporal membrane dynamics give rise to input-output relationships that are not fully observable through individual query-response interactions. Taken together with prior findings on data-centric privacy attacks, this work contributes to an emerging picture of SNNs as architecturally privacy-resilient models, not by design, but as a consequence of their computational substrate. Future work may explore attack strategies tailored to temporal representations and investigate how different neuron models and encoding schemes affect model extractability.
References
- [1]Anthony N Burkitt. 2006. A review of the integrate-and-fire neuron model: I. Homogeneous synaptic input. Biological Cybernetics 95, 1 (2006), 1–19.pubmed.ncbi.nlm.nih.gov/16622699
- [2]Niklas Bussmann, Paolo Giudici, Dimitri Marinelli, and Jochen Pabennbrok. 2021. Machine Learning for Financial Risk Management. Journal of Risk Finance (2021).
- [3]Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. 2017. ZOO: Zeroth Order Optimization Based Black-box Attacks to Deep Neural Networks without Training Substitute Models. In Proceedings of the ACM Conference on Computer and Communications Security (CCS).DOI
- [4]João Correia-Silva, Rodrigo F. Berriel, Claudine Badue, Alberto F. de Souza, and Thiago Oliveira-Santos. 2018. Copycat CNN: Stealing Knowledge by Persuading Confession with Random Non-Labeled Data. In Proceedings of the IEEE International Conference on Computer Vision (ICCV). 427–436.
- [5]Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. 2015. Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures. In Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 1322–1333.DOI
- [6]Wulfram Gerstner and Werner M Kistler. 2002. Spiking Neuron Models: Single Neurons, Populations, Plasticity. Cambridge University Press.
- [7]Matthew Jagielski, Nicholas Carlini, David Berthelot, Alex Kurakin, and Nicolas Papernot. 2020. High Accuracy and High Fidelity Extraction of Neural Networks. arXiv preprint arXiv:1909.01838 (2020).arxiv.org/abs/1909.01838
- [8]Mika Juuti, Sebastian Szyller, Samuel Marchal, and N Asokan. 2019. PRADA: protecting against DNN model stealing attacks. In 2019 IEEE European Symposium on Security and Privacy (EuroS&P). IEEE, 512–527.arxiv.org/abs/1805.02628
- [9]Alex Krizhevsky. 2009. Learning multiple layers of features from tiny images. Technical Report. University of Toronto.
- [10]Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient-based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278–2324.
- [11]Wolfgang Maass. 1997. Networks of spiking neurons: The third generation of neural network models. Neural Networks 10, 9 (1997), 1659–1671.DOI
- [12]Ayana Moshruba, Hsen Alouani, and Maryam Parsa. 2025. Are Neuromorphic Architectures Inherently Privacy-Preserving? An Exploratory Study. Proceedings on Privacy Enhancing Technologies (PoPETs) (2025).
- [13]Ayana Moshruba and Maryam Parsa. 2025. Edge-Private Spiking Graph Convolutional Learning: Membership Inference and Structural Differential Privacy. In 2025 First International Conference on Intelligent Computing and Systems at the Edge (ICEdge), Vol. 1–7. doi:10.1109/ICEdge67252.2025.11412839DOI
- [14]Ayana Moshruba, Hamed Poursiami, and Maryam Parsa. 2025. Izhikevich-Inspired Temporal Dynamics for Enhancing Privacy, Efficiency, and Transferability in Spiking Neural Networks. In 2025 International Conference on Neuromorphic Systems (ICONS). 89–97. doi:10.1109/ICONS60125.2025.00023arxiv.org/abs/2505.04034
- [15]Ayana Moshruba, Shay Bnyder, Hamed Poursiami, and Maryam Parsa. 2025. On the Privacy-Preserving Properties of Spiking Neural Networks with Unique Surrogate Gradients and Quantization Levels. In 2025 International Joint Conference on Neural Networks (IJCNN). 1–8. doi:10.1109/IJCNN64981.2025.11228352DOI
- [16]Emre O Neftci, Heshan Mostafa, and Friedemann Zenke. 2019. Surrogate Gradient Learning in Spiking Neural Networks. IEEE Signal Processing Magazine 36, 6 (2019), 61–63.
- [17]Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. 2019. Knockoff Nets: Stealing Functionality of Black-Box Models. In CVPR. 4954–4963.arxiv.org/abs/1812.02766
- [18]Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. 2017. Practical Black-Box Attacks against Machine Learning. In AsiaCCS.arxiv.org/abs/1602.02697
- [19]Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. 2016. Transferability in Machine Learning: from Phenomena to Black-Box Attacks using Adversarial Samples. In USENIX Security Symposium.arxiv.org/abs/1605.07277
- [20]Hamed Poursiami, Ihsen Alouani, and Maryam Parsa. 2024. BrainLeaks: On the Privacy-Preserving Properties of Neuromorphic Architectures against Model Inversion Attacks. In ICMLA.DOI
- [21]Hamed Poursiami, Ayana Moshruba, and Maryam Parsa. 2025. Do Spikes Protect Privacy? Investigating Black-Box Model Inversion Attacks in Spiking Neural Networks. arXiv preprint arXiv:2502.05509 (2025).DOI
- [22]W. Nicholson Price and I. Glenn Cohen. 2019. Privacy Risks of Machine Learning in Health Care. A Review. npj Digital Medicine 2, 1 (2019), 1–6.
- [23]Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. 2019. Towards Spike-Based Machine Intelligence with Neuromorphic Computing. Nature 575 (2019), 607–617.DOI
- [24]Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. 2017. Membership Inference Attacks Against Machine Learning Models. In 2017 IEEE Symposium on Security and Privacy (SP). IEEE, 3–18.arxiv.org/abs/1610.05820
- [25]Amirhossein Tavanaei, Masoud Ghodrati, Saeed Reza Kheradpisheh, Timothée Masquelier, and Anthony Maida. 2019. Deep Learning in Spiking Neural Networks. Neural Networks 111 (2019), 47–63.DOI
- [26]Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. 2018. Ensemble Adversarial Training: Attacks and Defenses. In International Conference on Learning Representations (ICLR).arxiv.org/abs/1705.07204
- [27]Florian Tramèr, Fan Zhang, Ari Juels, Michael K. Reiter, and Thomas Ristenpart. 2016. Stealing Machine Learning Models via Prediction APIs. In 25th USENIX Security Symposium (USENIX Security 16). 601–618.arxiv.org/abs/1609.02943
- [28]Han Xiao, Kashif Rasul, and Roland Vollgraf. 2017. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017).arxiv.org/abs/1708.07747
- [29]Samuel Yeom, Irene Giacomelli, Matt Fredrikson, and Somesh Jha. 2018. Privacy Risk in Machine Learning: Analyzing the Connection to Overfitting. IEEE CSF (2018).arxiv.org/abs/1709.01604
- [30]Zhi Zhou, Xuebin Chen, Enmin Li, Lianyong Zeng, Kai Luo, and Jun Zhang. 2019. Edge Intelligence: Paving the Last Mile of Artificial Intelligence with Edge Computing. Proc. IEEE 107, 8 (2019), 1738–1762.arxiv.org/abs/1905.10083






