VETO: Towards Protecting Images From Frontier AI Editing
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "VETO: Towards Protecting Images From Frontier AI Editing".
Paper summary: Tom: So, looking at this whole discussion about VETO: Towards Protecting Images From Frontier AI Editing Jonas Grebe et al., the authors are essentially saying they've built a way to stop those powerful new models from easily hijacking source images for harmful edits.
Jane: It boils down to Veto being an image cloak that specifically disrupts how these unified reference based editing systems work, especially when they are using joint attention blocks >
Lu: The authors use VetoBench to systematically check these defenses across different types of edits, from localized changes to recontextualizing objects in entirely new scenes >
Meng: For the people shipping models, what does this mean practically? Is this something we can just drop into our pipelines and expect it to work reliably without massive retraining?
Tom: They show that Veto substantially reduces edit success while keeping image quality high and it transfers across different model variants and downstream adaptations >
Jane: The authors are arguing that effective protection moving forward will require layered defenses, combining this image cloaking with inference level guardrails and parameter level guardrails >
Conclusion: Tom: So, we've been diving into Veto, this paper about protecting images from those super powerful new AI editing models called FLUX.two >
Jane: Right, it’s all about this system called Veto which acts like an image cloak to stop the AI from easily messing with your source pictures >
Tom: The authors are Jonas Grebe and his team, and they introduce a benchmark called VetoBench to test these defenses across different kinds of edits >
Jane: They're showing that this approach works better than what we had before because it looks at both small tweaks and bigger changes to the image context >
Tom: It seems like the main thing is that Veto manages to keep the picture looking good while stopping the AI from making big, unauthorized changes >
Jane: But what does this mean for us when we think about using these powerful generative models in real applications?
Tom: It means we have a method here that substantially cuts down on those edit successes across different tests, even when you look at the worst cases >
Jane: It suggests that layered defenses, combining this image cloaking with other safeguards in the system, is going to be necessary >
TU Darmstadt · Hessian.AI
cs.CV
Submitted: 2026-07-29
Updated: 2026-11-18
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: The rise of powerful, accessible image-editing models such as FLUX.2 has brought high-fidelity editing within broad reach and this expanded generative freedom also broadens the space of potential…
Key concepts
- Veto
- Veto is a user-controlled defense that acts as an 'anti-edit cloak' against Diffusion Transformer (DiT)-based editors. It works by maximizing a specific objective function to increase the entropy of reference attention, which disrupts the reliable extraction and preservation of information from the original source image during editing.
- Open-frame Edits
- These are a type of editing where an AI model extracts a subject's identity or concept from an image and then places that extracted element into a completely new, fabricated scene. VetoBench specifically tests protection against these more complex, context-shifting manipulations.
- Closed-frame Edits
- These are localized edits where the AI modifies only specific areas of an existing image while keeping the original scene intact. VetoBench distinguishes between these two types of edits to evaluate how well a defense protects both localized and broader contextual changes.
- Pareto Frontier Analysis
- This analysis uses a set of experiments to find the optimal balance between two competing goals: maximizing edit protection and maintaining high image fidelity (quality). The Pareto frontier shows the best possible trade-off points, demonstrating that Veto offers a superior trade-off compared to previous defenses.
Terminology
Summary
The rise of powerful, accessible image-editing models such as FLUX. p. 02 has brought high-fidelity editing within broad reach and this expanded generative freedom also broadens the space of potential misuse <ref:2607. p. 027292#pg2>. Veto introduces a subtle anti-edit cloak that disrupts the inner mechanism through which modern models read the source image <ref:2607. p. 027292#pg5> and VetoBench evaluates defenses not only on conventional localized edits but also on broader contextual shifts. p. 0 Across two contemporary editing models and three benchmarks, Veto consistently outperforms existing defenses while providing a stronger protection-fidelity trade-off. p. 0
Introduction to the Threat Landscape
Generative models have progressed from specialized text-to-image generation toward unified systems capable of both high-fidelity synthesis and instruction-based editing. This unified formulation enables edits beyond the source scene, extracting referenced entities and recontextualizing them in newly synthesized settings. p. 0 However, this extended creative freedom also introduces a severe security gap. p. 0 Obtaining convincing modifications requires specialized expertise as a single image can be exploited via natural-language instructions to place a real individual into a fabricated yet highly realistic scene. p. 0 Targeted visual misinformation, non-consensual sexual imagery, and other forms of reputational abuse are therefore no longer hypothetical concerns, but practical modes of misuse. p. 0
Mechanism of Veto Defense
Veto is a user-controlled, cost-raising defense against direct use of protected images in modern Diffusion Transformer (DiT)-based editors. By embedding a subtle anti-edit cloak, Veto substantially reduces edit success in standard open-weight pipelines. p. 1 The cloak maximizes the entropy of reference attention, disrupting the reliable extraction and preservation of source-image information. p. 1 This is achieved by maximizing the objective LVETO = H(A c->x) + H(A x->c) (1) over a configurable subset of flow steps, attention heads, and double-stream or single-stream layers. p. 3 Maximizing this objective flattens the canvas-to-source and source-to-canvas attention blocks. p. 3
Benchmark Design for Comprehensive Evaluation
VetoBench is the first benchmark for evaluating image protection beyond conventional edits. It distinguishes “closedframe” edits, which modify localized regions while preserving the original scene, from “open-frame” edits, which extract a subject’s identity or concept and recontextualize it within a newly generated scene. p. 1 Across 300 manually curated examples spanning general, defamatory, and graphic scenarios, VetoBench evaluates protection success using Multimodal Large Language Models (MLLMs). p. 1 This benchmark contains 300 examples evenly divided across general editing, defamation, and graphic violence. p. 3
Performance Against Baselines and Datasets
Across two models and three benchmarks, Veto consistently improves the trade-off between edit protection and image fidelity over prior defenses. Table 1 shows that Veto achieves the lowest residual human-judged edit success under the Pareto-selected perturbation budget, corresponding to the strongest protection. p. 4 VetoBench results show that for FLUX. p. 62, unprotected edit success is highest for general edits, reaching 91% under the MLLM judge and 94% under human evaluation. p. 4 Veto achieves a residual human-judged edit success of 1. p. 433% and 1.67% for FLUX. p. 02 under the Pareto-selected perturbation budget. p. 4
Robustness and Generalization
Veto retains the strongest protection-fidelity trade-off across all three domains, with ϵ = 4 reducing human-judged edit success on FLUX.2 to only 12 of 300 samples. p. 4 VetoBench results show that Veto remains the strongest defense in both regimes and achieves near-perfect protection on closed-frame edits. p. 5 Table 5 shows that stronger perturbations and targeted augmentation during optimization can partially mitigate horizontal flips and crops, whereas JPEG compression remains difficult under tight perturbation budgets. p. 5 Veto also transfers robustly across downstream model adaptations. p. 5
Ablation Study Insights
The ablation study indicates that the Canvas↔Ref interaction on the first double-stream block is most effective for FLUX.2, reducing MLLM edit success to 3. p. 533%. This aligns with the editing mechanism, since faithful generation depends on the canvas retrieving and preserving source content from the reference tokens. p. 5 The Canvas↔Ref + Ref↔Text configuration provides no further protection and slightly increases distortion. p. 6 The first double-stream block is selected as the default because it reduces edit success to 3. p. 633%, compared to 49. p. 467% for the singlestream hook at similar perceptibility. p. 6
Conclusion and Future Directions
We introduced Veto, an image cloak that disrupts unified reference-based editing, and VetoBench, a benchmark spanning closed- and open-frame manipulations to support systematic evaluation of modern anti-editing defenses. Across models and editing scenarios, Pareto-fair experiments show that Veto substantially reduces edit success while preserving image quality and transferring across model variants and downstream adaptations. p. 6 p. 6 Preliminary results in the supplementary material further indicate that Veto can protect individual sources in multi-reference editing settings. p. 6 Robustness to image transformations, architectural shifts, and future adaptation methods nevertheless remains open. p. 6 Effective protection will therefore require layered defenses that combine image cloaking with inference- and parameter-level guardrails. p. 6 p. 6
Supplementary Material Details
Appendix A first reviews the architecture and operation of modern unified image-editing models and explains how their design motivates the Veto objective. Appendix B then describes the computational infrastructure, experimental setup, and implementation of Veto and the evaluated baselines. p. 3 Appendix C presents the Pareto frontiers underlying the AnyEdit and EditBench results reported in the main paper, before Appendix D introduces the motivation, design, and construction of VetoBench. p. 4
Evaluation Protocol Details
We adopt Directional CLIP Similarity (CLIPdir) to measure whether the visual change aligns with the edit instruction. We use Multimodal Large Language Models (MLLMs) as human proxies to assess high-level edit success. p. 1 The MLLM is instructed to output a binary decision (YES or NO) based on the formal definitions of edit success and failure. p. 4 Human participants completed a short tutorial before proceeding, ensuring they understood the four failure modes in Figure 13a. p. 3
Qualitative Results Summary
Figure 18 demonstrates that Veto’s protection generalizes to higher image resolutions and different aspect ratios as well. Figure 15 shows additional qualitative results on VetoBench for FLUX. p. 52, where best viewed digitally with zoom. p. 5 Figure 18 also demonstrates that Veto successfully protects images at higher resolutions (shown here for up to 1024×1024 pixels) and various aspect ratios against FLUX. p. 42 editing. Figure 17 shows qualitative results on EditBench, AnyEdit, and VetoBench (General, Defamation, Gore) for Fibo-Edit. p. 6
Data Collection Methodology
We use an MLLM (Google Gemini 3.1 Pro) to generate candidate benchmark samples. p. 3 The MLLM is used only to propose and refine candidate scenarios. p. 3 All samples included in the final benchmark are manually selected and validated. p. 3 Two authors independently reviewed every candidate image-instruction pair. p. 0 Candidates were discarded if the requested edit was not applicable to the source image, or if the editing instruction was ambiguous or insufficiently specified. p. 0
Experimental Setup and Metrics
All experiments were conducted on a node running Ubuntu 22.04.4 LTS with an AMD EPYC 7313 16-Core Processor, 2. p. 50 TiB of system memory, and eight NVIDIA A100-SXM4-80GB GPUs. p. 8 We report Directional CLIP Similarity (CLIPdir) and MLLM-based Visual Question Answering (VQA) as primary quantitative metrics. p. 4 The Edit Success Rate (ESR) is defined based on four criteria: the requested edit is clearly present, no unrequested changes are introduced, original subject or object identities are preserved, and the image remains free of visual defects.
Pareto Frontier Analysis
Figure 9 complements these results by presenting the Pareto frontiers used to select the corresponding operating points. Across both benchmarks, Veto consistently provides a more favorable trade-off between protection effectiveness and perceptual similarity. p. 0 Fibo-Edit is particularly challenging for both baseline methods, whereas Veto remains effective at larger perturbation budgets. p. 5
Limitations of Existing Benchmarks
Existing image editing benchmarks exhibit several important limitations, including low-quality source images, physically impossible edit instructions, highly ambiguous prompts, and non-contextualized instructions. VetoBench introduces high-quality, context-rich evaluation tasks, including the novel openframe editing setting. p. 3 This balance is intentional and ensures equal coverage of the six combinations of application domain and edit type. p. 6
Final Conclusion
We introduced Veto, an image cloak that disrupts unified reference-based editing, and VetoBench, a benchmark spanning closed- and open-frame manipulations to support systematic evaluation of modern anti-editing defenses. Across models and editing scenarios, Pareto-fair experiments show that Veto substantially reduces edit success while preserving image quality and transferring across model variants and downstream adaptations. Effective protection will therefore require layered defenses that combine image cloaking with inference- and parameter-level guardrails.
Acknowledgements
The research was funded by a LOEWE-Spitzen-Professur (LOEWE/4a//519/05.00.002-(0010)/93) and has benefited from the Excellence Cluster “Reasonable AI” by the German Research Foundation (Deutsche Forschungsgemeinschaft - DFG) under Germany’s Excellence Strategy – EXC-3057. p. 7 Additionally, the research was partially funded by an Alexander von Humboldt Professorship in Multimodal Reliable AI, sponsored by the Federal Ministry of Research, Technology, and Space (BMFTR). p. 7 For compute, we gratefully acknowledge support from the hessian. p. 7AI Service Center (funded by the Federal Ministry of Research, Technology and Space (BMFTR), grant no. p. 7 16IS22091) and the hessian. p. 7AI Innovation Lab (funded by the Hessian Ministry for Digital Strategy and Innovation, grant no. p. 7 S-DIW04/0013/003). p. 7 Hossein Shakibania and Tobias Braun are supported by the Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA) through the DAAD program Konrad Zuse Schools of Excellence in Artificial Intelligence, sponsored by the German Federal Ministry of Education and Research. p. 7
References
Braun, T.; Grebe, J. p. 0 H.; Rohrbach, M. p. 7 p. 7; and Rohrbach, A. p. 7 p. 7 2026. Erased but Not Forgotten: How Backdoors Compromise Concept Erasure. p. 7 In Forty-third International Conference on Machine Learning. p. 7 p. 7 Braun, T.; Rothermel, M. p. 7; Rohrbach, M.; and Rohrbach, A. 2025. DEFAME: Dynamic Evidence-based FAct-checking with Multimodal Experts. p. 7 In Forty-second International Conference on Machine Learning. p. 7 Brooks, T.; Holynski, A. p. 0; and Efros, A. p. 0 A. 2023. Instructpix2pix: Learning to follow image editing instructions. p. 7 In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 18392–18402. p. 7 Chen, R.; Jin, H.; Liu, Y.; Chen, J.; Wang, H.; and Sun, L. p. 7 2024. EditShield: Protecting Unauthorized Image Editing by Instruction-guided Diffusion Models. p. 7 In European Conference on Computer Vision (ECCV). p. 7 Comanici, G. p. 4; Bieber, E. p. 7; Schaekermann, M. p. 7; Pasupat, I. p. 7; Sachdeva, N. p. 7; Dhillon, I. p. 7; Blistein, M. p. 7; Ram, O.; Zhang, D. p. 1; Rosen, E. p. 7, et al. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality. p. 7 Couairon, G. p. 0; Verbeek, J. p. 7; Schwenk, H. p. 7; and Cord, M. p. 7 2022. Diffedit: Diffusion-based semantic image editing with mask guidance. p. 7 Dong, Y.; Liao, F.; Pang, T.; Su, H.; Zhu, J.; Hu, X.; and Li, J. 2018. Boosting adversarial attacks with momentum. p. 7 In Proceedings of the IEEE conference on computer vision and pattern recognition, 9185–9193. p. 7 Esser, P.; Kulal, S. p. 7; Blattmann, A. p. 7; Entezari, R. p. 7; Müller, J. p. 7; Saini, H. p. 7; Levi, Y.; Lorenz, D. p. 7, Sauer, A. p. 7, Boesel, F. p. 7, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. p. 7 In Forty-first international conference on machine learning. p. 7 Gal, R.; Patashnik, O. p. 7; Maron, H. p. 7; Bermano, A. p. 7 H.; Chechik, G. p. 7, and Cohen-Or, D. p. 7 2022. Stylegan-nada: Clip-guided domain adaptation of image generators. p. 7 Gandikota, R. p. 0; Materzynska, J. p. 7; Fiotto-Kaufman, J. p. 7; and Bau, D. p. 7 2023. Erasing concepts from diffusion models. p. 7 In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2426–2436. p. 7 Gao, D.; Lu, S.; Walters, S. p. 7; Zhou, W.; Chu, J.; Zhang, J. p. 1; Zhang, B. p. 1, Jia, M., Zhao, J., Fan, Z., et al. 2025. EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers. p. 7 In International Conference on Machine Learning. p. 7 Goodfellow, I. p. 1 p. 1 J.; Pouget-Abadie, J. p. 7; Mirza, M. p. 7; Xu, B.; Warde-Farley, D. p. 7; Ozair, S. p. 7; Courville, A. p. 7, and Bengio, Y. p. 2 2014. Generative adversarial nets. p. 7 Advances in neural information processing systems, 27. p. 7 Goodfellow, I. J.; Shlens, J. p. 2, and Szegedy, C. p. 2 2014. Explaining and Harnessing Adversarial Examples. p. 7 Grebe, J. H.; Braun, T. p. 1; Rohrbach, A. p. 7; and Rohrbach, M. p. 7 2026. GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows. p. 7 In Forty-third International Conference on Machine Learning. Gui, D.; Guo, X.; Zhou,W.; and Lu, Y. 2025. I2VGuard: Safeguarding Images against Misuse in Diffusion-based Imageto-Video Models. p. 7 In Computer Vision and Pattern Recognition (CVPR). p. 7 Gutflaish, E. p. 1; Kachlon, E. p. 7; Zisman, H. p. 7; Hacham, T. p. 7; Sarid, N. p. 7; Visheratin, A. p. 7; Huberman, S. p. 7; Davidi, G. p. 7, Bukchin, G. p. 7, Goldberg, K. p. 7, et al. 2025. Generating an Image From 1000 Words: Enhancing Text-to-Image With Structured Captions. p. 7 Ho, J.; Jain, A.; and Abbeel, P. p. 1 2020. Denoising diffusion probabilistic models. p. 7 Advances in neural information processing systems, 33: 6840–6851. p. 7 Hosseini, H. p. 2; and Poovendran, R. p. 2 2018. Semantic Adversarial Examples. p. 7 In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 1614–1619. p. 7 Hu, E. J.; Shen, Y.; Wallis, P. p. 7; Allen-Zhu, Z. p. 7; Li, Y.; Wang, S.; Wang, L., and Chen, W. p. 1 2022. LoRA: Low-Rank Adaptation of Large Language Models. p. 7 In International Conference on Learning Representations. p. 7 p. 7 Huang, H.; Ma, X.; Erfani, S. p. 7 M.; Bailey, J. p. 7, and Wang, Y. p. 7 2021. Unlearnable Examples: Making Personal Data Unexploitable. p. 7 In International Conference on Learning Representations. Kawar, B.; Zada, S.; Lang, O.; Tov, O.; Chang, H. p. 7; Dekel, T. p. 7, Mosseri, I. p. 7, and Irani, M. p. 7
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Veto introduces a user-controlled, cost-raising defense that disrupts modern image editing models by embedding a subtle anti-edit cloak to maximize the entropy of reference attention, thereby disrupting
the reliable extraction and preservation of source-image information.
-
VetoBench provides a new benchmark that evaluates defenses not only on conventional localized edits but also on broader contextual shifts, distinguishing between
closedframe
edits which modify localized regions while preserving the original scene andopen-frame
edits which extract a subject’s identity or concept and recontextualize it within a newly generated scene. -
The improved system can achieve stronger protection-fidelity trade-offs across two contemporary editing models (FLUX.2 and Fibo-Edit) compared to existing defenses, consistently reducing edit success to as low as 1.33% for FLUX.2 under the Pareto-selected perturbation budget of epsilon equals 4.
-
Veto's effectiveness is demonstrated across various editing scenarios, including general editing, defamation (using synthetic subjects), and gore (including graphic violence), maintaining the strongest protection against unauthorized modifications in these domains across closed-frame and open-frame edits.
-
The system can generalize its protection to higher resolutions and different aspect ratios as shown in qualitative results on FLUX.2 up to 1024x1024 pixels, indicating robustness beyond the primary 512x512 experimental setting.
Sources
- Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
- DiffEdit: Diffusion-based semantic image editing with mask guidance
- Explaining and Harnessing Adversarial Examples
- Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions
- FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
- Flow Matching for Generative Modeling
- Towards Deep Learning Models Resistant to Adversarial Attacks
- Intriguing properties of neural networks
- GuardDoor: Safeguarding Against Malicious Diffusion Editing via Protective Backdoors
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs