WorksComputer Science
PREEMPT: Proactive Robustness Evaluation under Editing for Media Protection Techniques
Abstract
Proactive image protection methods aim to defend visual content against unauthorized manipulation and generative misuse by embedding protective signals into images before release. Evaluating whether these protections remain effective under image editing is therefore critical. However, existing evaluations remain largely limited to one-shot edits, providing only a static view of robustness while overlooking its evolution under realistic multi-step editing workflows. To systematically characterize robustness dynamics under sequential editing across various protection paradigms, we introduce PREEMPT, a holistic evaluation framework. It evaluates protection robustness under both one-shot and sequential editing scenarios, covering image transformations, semantic edits, homogeneous and heterogeneous edit compositions. To enable consistent analysis across protection paradigms with different underlying objectives, PREEMPT normalizes task-specific effectiveness into protection survival trajectories and summarizes long-horizon robustness through trajectory-based diagnostics. Using this framework, we reveal that degradation behaviors remain hidden under conventional one-shot evaluation. Furthermore, perturbation- and watermark-based protection methods exhibit different failure modes, and the first semantic edit largely determines subsequent degradation trajectories. PREEMPT provides a standardized protocol for studying long-horizon robustness and offers new insights into the degradation dynamics of proactive image protection.
Introduction
Recent advances in generative AI have made high-quality image editing increasingly accessible. Modern diffusion models [13, 38] enable realistic image manipulation [1, 6] and personalization [10, 14, 33] from only a few user instructions, but these capabilities also raise concerns about unauthorized content modification, identity misuse, and style imitation.
To mitigate such risks, proactive image protection methods [9, 16, 34, 43, 47] have been proposed to embed protective signals into images before release. Existing approaches mainly fall into two categories: perturbation-based methods [24, 36, 48], which attempt to disrupt downstream generation behaviors, and watermark-based methods [7, 22, 51], which enable ownership verification and content tracing after manipulation.
Existing robustness evaluations [11, 12, 49, 52] suffer from three major limitations. First, experimental protocols differ substantially across studies, with inconsistent choices of datasets, editing models, and evaluation criteria. Second, robustness is typically evaluated under isolated one-shot edits, while realistic image manipulations often involve multiple sequential operations whose cumulative effects remain poorly understood. Third, evaluations are method-specific: perturbation- and watermark-based methods are designed for different objectives and consequently rely on different performance metrics. Together, these limitations hinder fair comparison across protection methods and provide only a limited understanding of robustness under realistic edits.
To address these limitations, we introduce the Proactive Evaluation under Editing for Media Protection Techniques, (PREEMPT), a robustness evaluation framework for proactive image protection. PREEMPT systematically evaluates protection methods under both one-shot and long-horizon sequential editing scenarios, covering homogeneous semantic chains and heterogeneous mixed editing workflows. To enable consistent analysis across heterogeneous protection paradigms, PREEMPT further introduces a trajectory-based robustness representation consisting of Protection Survival Ratio (PSR), Protection Survival Curves, and Protection AUC, which respectively provide a unified robustness representation, trajectory-level visualization, and compact robustness summarization across heterogeneous protection paradigms. Our contributions are:
We propose PREEMPT, the first unified robustness evaluation framework for proactive image protection, which systematically evaluates protection methods under both
one-shot and long-horizon sequential editing scenarios.
We introduce a trajectory-based robustness characterization based on the Protection Survival Curves, together with Protection Survival Ratio (PSR) and Protection AUC, enabling consistent robustness analysis across perturbation- and watermark-based protection paradigms.
Using PREEMPT, we uncover several previously overlooked properties of proactive image protection, including continuous robustness evolution under sequential editing, paradigm-specific evolution patterns, ranking instability beyond one-shot evaluation, and the dominant role of semantic edits in long-horizon robustness.
Related Work
Proactive Image Protection
Proactive image protection aims to safeguard visual assets before they are released by embedding protective signals into images, thereby preventing unauthorized editing, personalization, or redistribution. Existing approaches can be broadly categorized into watermark-based and perturbation-based methods, which differ in both their protection objectives and underlying mechanisms.
Watermark-based methods [18, 22, 26, 53] embed identifiable information into images for ownership attribution, authenticity verification, and provenance tracing. Traditional watermarking techniques have recently been extended to diffusion-based generative models. Representative approaches include StegaStamp [40], which embeds recoverable signatures for copyright verification, Tree-Ring Watermarks [47], which inject structured patterns into diffusion latents for provenance tracing, Stable Signature [9], which integrates watermark generation into image synthesis, and WAM [35], which improves watermark robustness against common image editing and post-processing operations. Rather than preventing manipulation itself, these methods aim to preserve the recoverability of embedded messages after image modifications.
In contrast, perturbation-based methods [27, 36, 46, 50] proactively interfere with downstream image editing and model customization by introducing carefully optimized adversarial perturbations. Representative perturbation-based methods differ substantially in their protection mechanisms. PhotoGuard [34] disrupts the diffusion inference pipeline through adversarial perturbations, whereas Glaze [36] protects artistic styles by shifting feature representations toward surrogate targets. To disrupt the generation quality, AdvDM [21] further formulates perturbation generation directly using the diffusion objective, MIST [20] improves perturbation transferability via semantic-aware adversarial optimization. For personalization mimicry, Anti-DreamBooth [44] explicitly optimizes against the dynamic fine-tuning process. Unlike watermarking, perturbation-based methods seek to reduce the fidelity of downstream generation while remaining visually imperceptible.
Although the two paradigms pursue different protection objectives [41, 42], both fundamentally rely on the persistence of embedded protective signals throughout editing operations. Their robustness is therefore determined not only by the protection algorithm itself but also by the editing process that follows image release. This shared dependency motivates the need for a unified robustness evaluation framework capable of consistently assessing heterogeneous protection paradigms under realistic editing workflows.
Existing Robustness Evaluation
Existing robustness evaluation protocols differ across protection paradigms. Watermark-based methods [31] typically assess robustness by measuring message retrieval accuracy or decoding success rate after image manipulations, with the objective of preserving watermark detectability throughout the editing process. Consequently, robustness is commonly evaluated under a collection of image transformations or semantic editing attacks by examining whether the embedded message can still be recovered [26]. Perturbation-based methods usually evaluate two complementary aspects: perturbation imperceptibility and protection effectiveness. Imperceptibility is measured using perceptual image quality metrics such as LPIPS [52], while protection effectiveness is typically quantified by applying a single semantic edit and measuring the semantic similarity between the edited result and the protected image.
Besides semantic edits, robustness against image transformations has become a standard component of protection evaluation. Common operations include JPEG compression, Gaussian blur, crop-and-resize, horizontal flipping, and brightness/contrast adjustment, which are widely adopted to simulate the automatic preprocessing pipelines applied by online social networks (OSNs) after image upload [23]. Since adversarial perturbations are particularly vulnerable to these lossy transformations, many recent protection methods explicitly evaluate robustness under such operations to verify that the embedded perturbations remain effective after post-processing.
Despite adopting different objectives and evaluation metrics, existing protocols share several common limitations. First, robustness is predominantly evaluated under isolated one-shot manipulations. In contrast, real-world image editing is an iterative process. It typically involves sequential operations, continuous refinement, and history navigation [4, 5, 19]. Consequently, long-horizon robustness under sequential editing remains largely unexplored. Second, evaluations rely primarily on endpoint measurements after editing and therefore fail to characterize how protection robustness evolves throughout the editing process. Finally, the use of method-specific metrics prevents direct comparison across perturbation- and watermark-based protection paradigms. These limitations motivate the need for a unified robustness evaluation framework that not only enables fair comparison across heterogeneous protection paradigms but also characterizes robustness as an evolving process rather than a single endpoint.
PREEMPT
Evaluation Pipeline
To comprehensively simulate realistic post-deployment manipulations, we design a two-branch evaluation framework (see Figure 1) that evaluates protection robustness under both one-shot and long-horizon editing scenarios.
Figure 1. Overview of PREEMPT. Our framework evaluates proactive image protection methods under both one-shot and sequential editing scenarios. One-shot evaluation includes image-transformation-based edits and semantic edits, while sequential evaluation considers both homogeneous and heterogeneous edit chains to simulate realistic multi-step manipulations.
One-Shot Edit. PREEMPT first evaluates robustness under isolated editing operations. One-shot evaluation consists of two complementary categories: Image-Transformation-based and Semantic Edits. Image Transformations simulate common pixel-level image processing operations encountered after image release, whereas semantic edits evaluate robustness against content-level modifications generated by modern image editing models.
Transformation-based edits include commonly used image manipulations such as Gaussian blur, JPEG compression, crop-and-resize, horizontal flipping, and brightness/contrast adjustment. Semantic edits are organized into four categories: Content, Attribute, Style, and Spatial, covering modifications to semantic content, visual attributes, artistic style, and spatial configuration, respectively. Category-specific editing prompts are automatically generated by Vision-Language Models and executed by state-of-the-art image editing models to produce one-shot semantic manipulations. Together, transformation-based and semantic edits provide complementary evaluations of protection robustness under isolated editing scenarios.
Sequential Edit. PREEMPT further evaluates robustness under long-horizon editing workflows. Sequential edit pipelines are composed of multiple editing operation combinations. From homogeneous to heterogeneous settings, we consider two complementary settings: Simple Chains and Mixed Chains. Simple Chains consist of homogeneous semantic edits from a single semantic category, including content-only (C1), attribute-only (C2), style-only (C3), and spatial-only (C4). These chains isolate robustness degradation within individual semantic dimensions. Mixed Chains combine semantic edits with image-transformation operations to simulate practical editing workflows. Given a chain depth , edit operations are sequentially sampled from the editing pool to generate diverse editing trajectories.
Figure 2 presents representative heterogeneous editing trajectories. Each trajectory combines transformation-based and semantic edits to simulate realistic editing workflows, motivating the need for a unified representation of robustness evolution.
Protection Survival Representation
Existing proactive protection paradigms rely on different robustness metrics. Perturbation-based methods typically measure semantic similarity between the edited and original images, whereas watermark-based methods evaluate message retrieval accuracy. Since these paradigm-specific metrics differ in both scale and interpretation, directly comparing robustness across protection methods is difficult.
To quantify robustness evolution in a unified manner, we define the normalized Protection Survival Ratio (PSR) as:
where denotes the robustness metric adopted by the protection paradigm after editing depth , and is the metric value of the protected-but-unedited image. For perturbation-based methods, corresponds to semantic similarity, while for watermark-based methods it corresponds to message retrieval accuracy. By normalizing with respect to the initial protected state, PSR removes differences in metric scales and expresses robustness as a relative survival trajectory over editing depth. This formulation enables robustness evolution to be analyzed consistently across heterogeneous protection paradigms while preserving the interpretation of each underlying evaluation metric.
Notably, PSR is designed to characterize the evolution of robustness rather than to unify the underlying task-specific metrics themselves. Consequently, trajectory-based analyses, such as survival curves and Protection AUC, can be applied consistently to different protection paradigms without altering their original evaluation objectives. For perturbation-based methods under semantic edits, low similarity after editing can reflect either the intended semantic change or an additional effect of protection. To disentangle the two, we introduce a matched clean-trajectory control: unprotected images (“None”) undergo the same editing chains as protected images, and we report
where and are the paradigm-specific semantic similarities at depth . A positive indicates an additional protection-induced gap relative to the same clean editing path. Watermarking methods are evaluated with bit accuracy and are excluded from this control, as they do not share the same semantic-similarity confound.
Robustness Characterization
To characterize long-horizon robustness, PREEMPT provides both a visual representation and a compact summary of protection survival trajectories.
Protection Survival Curve. Given the normalized PSR, robustness under sequential editing can be represented as a protection survival curve, , where each point corresponds to the normalized protection effectiveness after editing depth . Starting from the normalized initial state (), the resulting curve characterizes how protection effectiveness evolves throughout sequential editing. The curve provides an intuitive representation of robustness evolution and facilitates comparison across protection methods and editing scenarios.
Protection AUC. While survival curves provide an intuitive visualization of robustness evolution, comparing massive trajectories across heterogeneous methods and scenarios can be cumbersome. We therefore summarize each trajectory using the area under the protection survival curve:
where denotes the maximum editing depth. Since , the resulting AUC is also normalized to , with larger values indicating stronger overall robustness throughout the editing process.
One-shot Robustness
Experimental Setup
To demonstrate PREEMPT, we instantiate the framework using representative proactive image protection methods from both perturbation-based and watermark-based paradigms. Specifically, we evaluate four perturbation-based methods (PhotoGuard [34], Mist [20], AdvDM [21], and SDST [50]) together with four recent watermarking methods (Robust-Wide [15], TrustMark [2], VINE [26] and WAM [35]). These methods are selected to demonstrate the applicability of PREEMPT across various protection paradigms. To ensure fair comparison, all perturbation-based methods are generated using a standard perturbation budget () and 100 optimization steps. Clean images serve as the no-protection baseline (None).
Protection effectiveness is evaluated using the task-specific metrics corresponding to each protection paradigm. For perturbation-based methods, semantic similarity is measured using domain-specific feature representations: ArcFace [8] identity similarity for biometric faces, CSD [37] style embedding similarity for artworks, and DINOv2 [30] feature similarity for natural scene images. The resulting scores are normalized to a common scale for cross-domain comparison. For watermark-based methods, robustness is measured by message retrieval bit accuracy using each method’s official encoding and decoding pipeline. Following the original implementations, we embed the message “Watermark” for VINE and “WAMK” for WAM according to their payload capacities, a fixed 64-bit binary payload for Robust-Wide, and a fixed 100-bit binary payload for TrustMark (ECC disabled).
Image-transformation-based edits include Gaussian blur, JPEG compression, crop-and-resize, horizontal flipping, and brightness/contrast adjustment. These transformations [45] represent typical post-processing encountered after image release. For semantic edits, four editing prompts are generated for each semantic category using GPT-4.1-mini [29] and executed by two representative image editing models, Stable Diffusion v2.1 [32] and InstructPix2Pix [1]. Since PREEMPT is independent of the underlying editing model, alternative editors can be incorporated without modifying the evaluation protocol.
Figure 2. Representative heterogeneous editing trajectories. Rows correspond to faces, artworks, and objects. Each trajectory consists of alternating transformation-based and semantic edits, illustrating the mixed editing workflow used in our sequential robustness evaluation.
The evaluation benchmark covers three representative visual domains: artwork images from WikiArt [39], biometric face images from CelebA-HQ [17, 25], and object-centric images from DreamBooth [33]. Curated subsets from these datasets are used throughout all experiments. Complete implementation details, parameter settings, and prompt templates are provided in the Supplementary Material (SM).
Image-Transformation-based edits
Image-transformation-based edits evaluate protection robustness under common post-processing operations encountered during image sharing. The quantitative results for perturbation- and watermark-based methods are reported in Table 1 and Table 2, respectively. We also report the “No edit” column as the no-transformation anchor (protected vs. clean before any transform).
| Method | No edit | JPEG | JPEG | Bright. & Contr. | Blur | Crop | Flip | Avg. |
| None | 1.000 | 0.988 | 0.975 | 0.976 | 0.971 | 0.940 | 0.966 | 0.969 |
| AdvDM | 0.872 | 0.874 | 0.872 | 0.869 | 0.880 | 0.824 | 0.858 | 0.863 |
| MIST | 0.901 | 0.895 | 0.889 | 0.890 | 0.881 | 0.855 | 0.884 | 0.882 |
| PhotoGuard | 0.900 | 0.895 | 0.890 | 0.890 | 0.880 | 0.856 | 0.885 | 0.883 |
| SDST | 0.909 | 0.899 | 0.896 | 0.896 | 0.884 | 0.865 | 0.890 | 0.888 |
Table 1. Effectiveness of perturbation protection methods under one-shot image transformation. Metric: semantic similarity (↓).
| Method | No edit | JPEG | JPEG | Bright. & Contr. | Blur | Crop | Flip | Avg. |
| Robust-Wide | 1.000 | 1.000 | 1.000 | 0.987 | 1.000 | 0.539 | 0.452 | 0.830 |
| TrustMark | 1.000 | 0.999 | 0.998 | 0.982 | 1.000 | 1.000 | 0.999 | 0.996 |
| VINE | 1.000 | 1.000 | 1.000 | 0.998 | 0.999 | 0.504 | 0.465 | 0.828 |
| WAM | 1.000 | 0.998 | 0.994 | 1.000 | 1.000 | 1.000 | 1.000 | 0.999 |
Table 2. Effectiveness of watermarking methods under one-shot image transformations. Metric: message retrieval accuracy (↑).
Both protection paradigms remain largely robust under common image transformations. Compared with the unprotected baseline, perturbation-based methods consistently reduce semantic similarity across all evaluated transformations, while watermark-based methods maintain nearly perfect message retrieval accuracy under most operations. Crop-and-resize produces the largest performance degradation for both paradigms, whereas JPEG compression, Gaussian blur, and brightness/contrast adjustment have comparatively minor effects. These results indicate that routine post-processing alone is generally insufficient to invalidate existing proactive image protection methods.
Semantic Edit
Unlike image transformations, semantic edits directly modify image content and therefore provide a more challenging evaluation scenario. Unless otherwise specified, we report results obtained using InstructPix2Pix [1] in the main paper, while the corresponding results using Stable Diffusion v2.1 [32] are provided in SM. The quantitative results are summarized in Table 3 and Table 4.
| Method | No edit | Content | Attribute | Style | Spatial | Avg. |
| None | 1.000 | 0.760 | 0.720 | 0.740 | 0.695 | 0.729 |
| AdvDM | 0.872 | 0.267 | 0.234 | 0.331 | 0.249 | 0.270 |
| MIST | 0.901 | 0.505 | 0.384 | 0.495 | 0.492 | 0.469 |
| PhotoGuard | 0.900 | 0.490 | 0.372 | 0.486 | 0.462 | 0.453 |
| SDST | 0.909 | 0.458 | 0.376 | 0.508 | 0.495 | 0.459 |
Table 3. Protection effectiveness of perturbation-based methods under one-shot semantic edits. Metric: semantic similarity (↓).
| Method | No edit | Content | Attribute | Style | Spatial | Average |
| Robust-Wide | 1.000 | 0.990 | 0.970 | 0.950 | 0.975 | 0.971 |
| TrustMark | 1.000 | 0.899 | 0.866 | 0.859 | 0.852 | 0.869 |
| VINE | 1.000 | 0.945 | 0.899 | 0.850 | 0.917 | 0.903 |
| WAM | 1.000 | 0.894 | 0.805 | 0.733 | 0.837 | 0.817 |
Table 4. Protection effectiveness of watermarking methods under one-shot semantic edits. Metric: message retrieval accuracy (↑).
Compared with image transformations, semantic edits pose a substantially greater challenge to proactive image protection. Perturbation-based methods exhibit markedly lower protection effectiveness after a single semantic manipulation, while watermark-based methods also experience measurable retrieval degradation despite maintaining relatively high decoding accuracy. Across both protection paradigms, content-level modifications have a considerably stronger impact than low-level image transformations.
Nevertheless, these one-shot evaluations only measure protection effectiveness after an isolated edit and therefore provide static endpoint measurements. Whether such observations remain valid under realistic multi-step editing workflows is investigated in the next section.
Long-Horizon Robustness
Experimental Setup
For long-horizon evaluation, we construct sequential editing chains with a default depth throughout the main paper. Results of Stable Diffusion v2.1, along with additional long-horizon experiments including sensitivity to chain length and edit ordering, are presented in SM.
We consider two complementary sequential settings. Simple chains repeatedly apply semantic edits from a single category, enabling controlled evaluation of robustness evolution within individual semantic dimensions. Complex chains consist of heterogeneous edit sequences by combining semantic edits and image transformations, simulating practical multi-step editing workflows. Unless otherwise specified, edit operations are randomly sampled using a fixed seed of 42 to ensure reproducibility.
We do not consider transformation-only chains. Consecutive image transformations primarily accumulate low-level distortions and can often be approximated by a single transformation with adjusted parameters. In contrast, sequential semantic edits progressively reconstruct image content, resulting in substantially different robustness evolution. We therefore focus homogeneous chains on semantic edits while using heterogeneous chains to approximate realistic editing workflows.
Robustness Evolution
Figure Figure 3 visualizes the evolution of protection robustness under sequential editing using the proposed PSR. Homogeneous semantic chains and heterogeneous mixed chains are shown separately to distinguish controlled semantic degradation from practical multi-step editing workflows. Across both sequential settings, robustness evolves continuously with editing depth rather than collapsing after a single manipulation. All protection methods exhibit progressively decreasing PSR as additional edits are applied, indicating that robustness should be regarded as a dynamic property instead of a static endpoint. Under the matched clean trajectory, unprotected images already exhibit substantial semantic drift—e.g., raw concept similarity drops to on homogeneous chains and on heterogeneous chains—showing that sequential semantic editing itself reduces the measured signal. Relative to this baseline, the evaluated perturbation methods consistently yield . Protection pushes similarity down by an additional gap beyond the intended edit (see SM for clean-trajectory curves and tables).
Figure 3. Protection survival curves under homogeneous semantic chains (left) and heterogeneous mixed chains (right). Robustness evolves progressively with editing depth rather than remaining a static endpoint. The x-axis denotes the editing depth (), and the y-axis shows the normalized Protection Survival Ratio (PSR). Solid lines indicate perturbation-based methods, and dashed lines indicate watermark-based methods. The legend reports each method together with its pre-edit effectiveness, .
This observation highlights an important limitation of conventional one-shot evaluation. Although measuring protection effectiveness after a single edit provides a useful snapshot of robustness, it does not reveal how protection evolves under realistic multi-step editing. Our results show that protection degrades progressively as edits accumulate, motivating trajectory-based analyses that capture robustness dynamics over the entire editing process.
Paradigm-specific Evolution
Figure Figure 4 compares the normalized robustness trajectories of perturbation- and watermark-based methods. Although both paradigms exhibit continuous robustness evolution under sequential editing, their evolution patterns differ substantially. Perturbation-based methods show a rapid change immediately after the first edit, followed by progressively diverging trajectories across methods. In contrast, watermark-based methods exhibit smoother and more consistent robustness evolution throughout the editing process. Sequential editing also increases the variability of robustness trajectories among perturbation-based methods, as reflected by the progressively wider spread across methods.
Figure 4. Protection survival curves aggregated by protection paradigms. Shaded regions indicate the standard deviation across methods within each protection paradigm. Perturbation- and watermark-based methods exhibit distinct robustness evolution patterns under sequential editing.
The observed difference is consistent with the different design objectives of the two protection paradigms. Perturbation-based methods protect images by disrupting downstream image generation, making their robustness closely coupled with semantic reconstruction during editing. Consequently, repeated semantic edits produce increasingly diverse robustness trajectories across different perturbation methods. In contrast, watermark-based methods aim to preserve an embedded signal for subsequent decoding. As long as sufficient watermark information survives each editing step, message retrieval remains relatively stable, leading to smoother robustness trajectories. These results indicate that robustness evolution is paradigm-dependent. Rather than enforcing a single robustness metric, PREEMPT accommodates paradigm-specific measurements within a unified trajectory-based evaluation framework, enabling consistent analysis and comparison of robustness evolution across different protection paradigms.
Ranking Stability
Ranking instability is primarily observed for perturbation-based methods. We focus the analysis on perturbation-based methods, where ranking changes are more informative. Figure 5 summarizes average ranking change from one-shot () to long-horizon () evaluation. Positive values indicate that a method drops in rank after sequential editing, whereas negative values indicate an improved ranking. Although all methods are evaluated under the same protocol, their rankings are not consistently preserved. For example, MIST, which ranks among the strongest methods after a single edit, consistently loses ranking under sequential evaluation, whereas AdvDM and SDST improve their ranks across multiple editing settings.
Figure 5. Average rank change from one-shot () to long-horizon () evaluation under homogeneous and heterogeneous edit chains. Positive (negative) values indicate worse (better) rankings after sequential editing, showing that one-shot rankings are not consistently preserved.
These observations suggest that one-shot robustness is not always predictive of long-horizon robustness. To quantify this gap beyond ranking changes, we test whether one-shot semantic robustness predicts long-horizon survival. For each protection method, we take the mean one-shot concept similarity over the four semantic edit types as the predictor, and use the normalized protection survival AUC (PAUC) under sequential editing as the target. Fitting a linear regression at the methodchain level yields only a moderate association (, ), indicating that one-shot scores explain little of the variance in long-horizon robustness. Thus, endpoint measurements after a single edit are insufficient to characterize protection under multi-step editing; trajectory-based evaluation remains necessary. Detailed scatter plots and correlation statistics are provided in SM. Methods exhibiting similar performance after a single edit may follow substantially different robustness trajectories under continued editing, eventually leading to different relative rankings. Consequently, evaluating proactive protection methods solely under one-shot settings may produce method rankings that do not generalize to realistic multi-step editing scenarios. Sequential evaluation therefore provides complementary information that cannot be inferred from one-shot robustness alone.
The Role of the First Semantic Edit
Having established that protection robustness evolves throughout sequential editing, we next investigate what drives this evolution. Since heterogeneous edit chains consist of both image transformations and semantic edits, we group editing trajectories according to whether the first operation is a physical transformation or a semantic edit.
Figure 6 compares the resulting robustness trajectories. When the first operation is an image transformation, protection robustness remains nearly unchanged after the first editing step, and the major robustness transition occurs only after subsequent semantic edits are introduced (see Figure 6 (a)). In contrast, when the first operation is semantic, robustness exhibits an immediate transition after the first editing step, followed by substantially slower evolution throughout the remaining editing process (see Figure 6 (b)). This contrast suggests that robustness evolution is more strongly associated with semantic edits than with low-level image transformations. Unlike image transformations, which mainly alter low-level image statistics while preserving semantic content, semantic edits explicitly reconstruct the image content, directly affecting the information that proactive protection methods seek to preserve or disrupt.
Figure 6. Robustness trajectories under heterogeneous edit chains grouped by the type of the first edit. (a) Physical-first trajectories remain largely unchanged until the first semantic edit occurs. (b) Semantic-first trajectories exhibit an immediate robustness transition followed by slower evolution. Shaded regions denote the standard deviation across perturbation-based methods.
We further define the first-step contribution:
which measures the fraction of the total robustness degradation occurring after the first editing step. The results are summarized in Table 5. Across homogeneous semantic chains, the first semantic edit accounts for approximately 63%–93% of the total degradation, indicating that most robustness change is concentrated in the initial semantic modification. In heterogeneous edit chains, however, this proportion consistently decreases to 44%–52%, suggesting robustness degrades more gradually across successive editing operations rather than being dominated by the first edit.
| Method | Homo. | Hetero. |
| PhotoGuard | 0.65 | 0.45 |
| MIST | 0.63 | 0.45 |
| AdvDM | 0.93 | 0.52 |
| SDST | 0.69 | 0.44 |
Table 5. Contribution of the first editing step to the overall robustness evolution. Higher indicates that a larger proportion of robustness change occurs immediately after the first edit.
Taken together, these results indicate that semantic edits serve as the primary transition governing perturbation robustness, whereas image transformations largely preserve the existing robustness state until semantic reconstruction occurs. More importantly, they demonstrate that, under the evaluated editing scenarios, the evolution of protection robustness depends not only on editing depth but also on the composition of the editing sequence.
Robustness under Purification Attacks
Another important category of post-processing operations that can impact robustness of protection includes regeneration or purification-style methods (e.g., DiffPure [28] and IMPRESS [3]). These are different from common transformation-based and semantic-based edits which are typically applied without assuming knowledge of the protection. Purification is performed by informed adversaries who explicitly attempt to remove the protective signal before further use. Although such stress tests are informative, they answer a different question: whether protection survives targeted removal. PREEMPT instead asks how long protection remains effective under a realistic distribution and editing lifecycle after release. Results evaluating resistance to targeted purification (DiffPure and IMPRESS) are included in SM for completeness.
Conclusion
In this paper, we present PREEMPT, a holistic evaluation framework for assessing the robustness of proactive image protection methods under realistic editing scenarios. Rather than attempting to enumerate all possible editing chains, PREEMPT evaluates robustness through representative one-shot and sequential editing scenarios that provide a practical and reproducible evaluation protocol. We introduce unified trajectory-based metrics to characterize robustness evolution across different protection paradigms. Beyond evaluating robustness, PREEMPT reveals three key insights. First, robustness evolves differently across protection paradigms. Second, similar one-shot robustness does not necessarily imply comparable long-horizon robustness, as methods may follow different degradation trajectories under sequential editing. Third, robustness degradation is largely concentrated in the initial semantic edit, highlighting its dominant role in long-horizon robustness evolution. These insights suggest that protection methods should optimize robustness over editing trajectories rather than a single operating point. We hope that PREEMPT provides a foundation for future proactive image protection research.
References
- [1]Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 18392–18402, 2023. 1, 4, 5arxiv.org/abs/2211.09800
- [2]Tu Bui, Shruti Agarwal, and John Collomosse. Trustmark: Robust watermarking and watermark removal for arbitrary resolution images. In IEEE International Conference on Computer Vision (ICCV), 2025. 4
- [3]Bochuan Cao, Changjiang Li, Ting Wang, Jinyuan Jia, Bo Li, and Jinghui Chen. IMPRESS: Evaluating the resilience of imperceptible perturbations against unauthorized data usage in diffusion-based generative AI. In Advances in Neural Information Processing Systems, 2023. 8arxiv.org/abs/2310.19248
- [4]Hsiang-Ting Chen, Li-Yi Wei, Björn Hartmann, and Maneesh Agrawala. Data-driven adaptive history for image editing. In Proceedings of the 20th ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, page 103–111, New York, NY, USA, 2016. Association for Computing Machinery. 2DOI
- [5]Tianyu Chen, Yasi Zhang, Zhi Zhang, Peiyu Yu, Shu Wang, Zhendong Wang, Kevin Lin, Xiaofei Wang, Zhengyuan Yang, Linjie Li, et al. Edival-agent: An object-centric framework for automated, fine-grained evaluation of multi-turn editing. arXiv preprint arXiv:2509.13399, 2025. 2arxiv.org/abs/2509.13399
- [6]Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. DiffEdit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1arxiv.org/abs/2210.11427
- [7]Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for data copyright protection against generative diffusion models. SIGKDD Explor. Newsl., 26(2):60–75, 2025. 1
- [8]Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690–4699, 2019. 4DOI
- [9]Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 22466–22477, 2023. 1, 2arxiv.org/abs/2303.15435
- [10]Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 1arxiv.org/abs/2208.01618
- [11]Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Conference on Empirical Methods in Natural Language Processing, pages 7514–7528. Association for Computational Linguistics, 2021. 1
- [12]Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017.
- [13]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020.arxiv.org/abs/2006.11239
- [14]Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 1arxiv.org/abs/2106.09685
- [15]Runyi Hu, Jie Zhang, Ting Xu, Jiwei Li, and Tianwei Zhang. Robust-wide: Robust watermarking against instruction-driven image editing. In European Conference on Computer Vision, pages 20–37. Springer, 2025. 4arxiv.org/abs/2402.12688
- [16]Jaehwan Jeong, Sumin In, Sieun Kim, Hannie Shin, Jongheon Jeong, Sang Ho Yoo, Jaewook Chung, and Sang-pil Kim. Faceshield: Defending facial image against deepfake threats. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10364–10374, 2025. 1arxiv.org/abs/2412.09921
- [17]Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations (ICLR), 2018. 5arxiv.org/abs/1710.10196
- [18]Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model, 2024. 2arxiv.org/abs/2405.02696
- [19]Hongyu Li, Manyuan Zhang, Dian Zheng, Ziyu Guo, Yimeng Jia, Kaituo Feng, Hao Yu, Yexin Liu, Yan Feng, Peng Pei, et al. Editthinker: Unlocking iterative reasoning for any image editor. arXiv preprint arXiv:2512.05965, 2025. 2arxiv.org/abs/2512.05965
- [20]Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2, 4arxiv.org/abs/2305.12683
- [21]Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning (ICML), pages 20763–20786. PMLR, 2023. 2, 4arxiv.org/abs/2302.04578
- [22]Gaozhi Liu, Siwu Cao, Zhenxing Qian, Xinpeng Zhang, Sheng Li, and Wanli Peng. Watermarking one for all: A robust watermarking scheme against partial image theft. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8225–8234, 2025. 1, 2DOI
- [23]Jun Liu, Jiantao Zhou, Haiwei Wu, Weiwei Sun, and Jinyu Tian. Generating robust adversarial examples against online social networks (osns). ACM Transactions on Multimedia Computing, Communications and Applications, 20(4):1–26, 2023. 2arxiv.org/abs/2310.12708
- [24]Yixin Liu, Chenrui Fan, Yutong Dai, Xun Chen, Pan Zhou, and Lichao Sun. Metacloak: Preventing unauthorized subject-driven text-to-image diffusion-based synthesis via meta-learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24219–24228, 2024. 1arxiv.org/abs/2311.13127
- [25]Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoo Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730–3738, 2015. 5
- [26]Shilin Lu, Zihan Zhou, Jiayou Lu, Yuanzhi Zhu, and Adams Wai-Kin Kong. Robust watermarking using generative priors against image editing: From benchmarking to advances. In The Thirteenth International Conference on Learning Representations, 2025. 2, 4
- [27]Xiaoyue Mi, Fan Tang, Juan Cao, Peng Li, and Yang Liu. Visual-friendly concept protection via selective adversarial perturbations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 39033–39041, 2024. 2arxiv.org/abs/2408.08518
- [28]Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning (ICML), 2022. 8arxiv.org/abs/2205.07460
- [29]OpenAI. Gpt-4 technical report, 2023. 4arxiv.org/abs/2303.08774
- [30]Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4arxiv.org/abs/2304.07193
- [31]Tahseen Rabbani, Bang An, Mucong Ding, Aakriti Agrawal, Yuancheng Xu, Chenghao Deng, Sicheng Zhu, Abdirisak Mohamed, Yuxin Wen, Tom Goldstein, and Furong Huang. WAVES: Benchmarking the robustness of image watermarks. In Privacy Regulation and Protection in Machine Learning, 2024. 2arxiv.org/abs/2401.08573
- [32]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 ieee. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4, 5arxiv.org/abs/2112.10752
- [33]Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500–22510, 2023. 1, 5arxiv.org/abs/2208.12242
- [34]Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious AI-powered image editing. In Proceedings of the 40th International Conference on Machine Learning, pages 29894–29918. PMLR, 2023. 1, 2, 4arxiv.org/abs/2302.06588
- [35]Tom Sander, Pierre Fernandez, Alain Dujmous, Teddy Furon, and Matthias Douze. Watermark anything with localized messages. In International Conference on Learning Representations (ICLR), 2025. 2, 4arxiv.org/abs/2411.07231
- [36]Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by text-to-image models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1, 2arxiv.org/abs/2302.04222
- [37]Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. Proceedings of the European conference on computer vision (ECCV), 2024. 4arxiv.org/abs/2404.01292
- [38]Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 1arxiv.org/abs/2010.02502
- [39]Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394–409, 2019. 5arxiv.org/abs/1708.09533
- [40]Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2arxiv.org/abs/1904.05343
- [41]Qiuyu Tang and Aparna Bharati. Watermarks vs. perturbations for preventing AI-based style editing. In The 1st Workshop on GenAI Watermarking, 2025. 2
- [42]Qiuyu Tang, Bonor Ayambem, Mooi Choo Chuah, and Aparna Bharati. Is perturbation-based image protection disruptive to image editing? 2025 IEEE International Conference on Image Processing (ICIP), 2025. 2arxiv.org/abs/2506.04394
- [43]Qiuyu Tang, Joshua Krinsky, and Aparna Bharati. Styleprotect: Safeguarding artistic identity in fine-tuned diffusion models. arXiv preprint arXiv:2509.13711, 2025. 1arxiv.org/abs/2509.13711
- [44]Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2116–2127, 2023. 2arxiv.org/abs/2303.15433
- [45]Luisa Verdoliva. Media forensics and deepfakes: an overview. IEEE journal of selected topics in signal processing, 14(5):910–932, 2020. 4DOI
- [46]Hanhui Wang, Yihua Zhang, Ruizheng Bai, Yue Zhao, Sijia Liu, and Zhengzhong Tu. Edit away and my face will not stay: Personal biometric defense against malicious generative editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2
- [47]Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2DOI
- [48]Jingyao Xu, Yuetong Lu, Yandong Li, Siyang Lu, Dongdong Wang, and Xiang Wei. Perturbing attention gives you more bang for the buck: Subtle imaging perturbations that efficiently fool customized diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24534–24543, 2024. 1DOI
- [49]Zitong Xu, Huiyu Duan, Bingnan Liu, Guangji Ma, Jiarui Wang, Liu Yang, Shiqi Gao, Xiaoyu Wang, Jia Wang, Xiongkuo Min, et al. Lmm4edit: Benchmarking and evaluating multimodal image editing with lmms. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 6908–6917, 2025. 1arxiv.org/abs/2507.16193
- [50]Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In International Conference on Learning Representations (ICLR), 2024. 2, 4arxiv.org/abs/2311.12832
- [51]Xiaoyu Ye, Hao Huang, Jiaqi An, and Yongtao Wang. DUAW: Data-free universal adversarial watermark against stable diffusion customization. In ICLR 2024 Workshop on Secure and Trustworthy Large Language Models, 2024. 1arxiv.org/abs/2308.09889
- [52]Richard Zhang, Phillip Isola, Alexei A Efros, Eli Schechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586–595, 2018. 1, 2arxiv.org/abs/1801.03924
- [53]Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11964–11974, 2024. 2arxiv.org/abs/2312.08883





