TL;DR
Modern unified image-editing models like FLUX.2 easily bypass legacy image-protection cloaks because they process text and images together through joint-attention mechanisms rather than isolated encoders. This paper introduces VETO, an adversarial cloak that protects images by maximizing the entropy of these internal attention pathways, alongside VetoBench, a new benchmark designed to evaluate image defenses against malicious, open-frame recontextualizations.
What Problem Does This Paper Address?
The rapid rise of Diffusion Transformer (DiT) models, such as FLUX.2, has fundamentally changed image editing. Older pipelines typically inverted a source image into a latent code and modified it via text prompts. Modern unified models, however, treat editing as reference-conditioned generation, allowing text prompts and generation tokens to attend directly to the source image tokens throughout the entire denoising process.
While this grants incredible creative freedom—enabling "open-frame" edits that extract a subject and place them in an entirely new scene—it also introduces a massive security gap. Attackers can easily generate non-consensual imagery or defamatory content using completely benign prompts (e.g., placing a person in a police lineup). Traditional safety filters fail here because neither the prompt nor the background is explicitly harmful in isolation.
Furthermore, existing image-cloaking defenses (like PhotoGuard or EditShield) target only the initial image encoder bottleneck. Because modern unified editors rely on strong generative priors and joint-attention mechanisms, they easily bypass these legacy protections.
Methods
To address this vulnerability, the authors present VETO (Anti-Edit Protection), a user-controlled adversarial cloak that adds a subtle, imperceptible pixel perturbation ($\delta$) to the source image.
Instead of attacking the initial image encoder, VETO directly targets the internal interaction between the source image tokens ($x$) and the evolving output canvas tokens ($c_t$) within the model's Multimodal Diffusion Transformer (MMDiT) blocks.
VETO uses a momentum iterative fast-gradient sign method (MI-FGSM) to optimize the perturbation by maximizing the blockwise entropy of the bidirectional source-canvas attention maps:
$$\mathcal{L}_{\text{VETO}} = H(A^{c \to x}) + H(A^{x \to c})$$
By maximizing this entropy, VETO flattens the attention distribution. This diffuses spatial attention, effectively preventing the model from reliably retrieving or transferring localized features from the protected source image to the final generated output.
Datasets
The authors evaluated their approach on three distinct datasets:
- EditBench: 300 samples focusing on standard, localized "closed-frame" edits (e.g., changing backgrounds, style, or weather).
- AnyEdit: 300 samples representing general closed-frame modifications that preserve the source composition.
- VetoBench (Newly Introduced): A curated 300-sample benchmark specifically designed to evaluate anti-editing protections against both closed-frame and complex open-frame recontextualizations. It is evenly split across three domains: General, Defamation, and Gore. To ensure ethical safety, all depicted individuals in VetoBench are entirely synthetic.
Key Insights and Ideas
- The Death of Encoder-Level Defenses: Traditional cloaks fail against modern DiT editors because displacing the initial latent encoding is an indirect proxy. The model's internal joint-attention layers can easily reconstruct or work around the perturbed encoding using its powerful generative priors.
- Targeting Attention Entropy: Rather than trying to push an image toward a specific "broken" state, forcing the cross-attention maps to be uniform (high entropy) is a highly effective way to disrupt information transfer without visibly degrading the original image.
- Contextual Harm Requires Image-Side Control: Because malicious edits often rely on benign prompts to create harmful contexts, centralized model-side filters are fundamentally ill-equipped. Defense must remain under the image owner's control via localized cloaking.
Key Results
VETO consistently achieved a superior trade-off between edit protection and image fidelity compared to prior defenses across multiple models (FLUX.2 and Fibo-Edit).
- EditBench Performance: On FLUX.2, VETO reduced the human-judged Edit Success Rate (ESR) from 66.00% (unprotected) down to just 1.33%. In comparison, PhotoGuard and EditShield only reduced it to 18.33% and 20.33%, respectively, while causing significantly higher perceptual distortion (measured via LPIPS).
- VetoBench Performance: VETO demonstrated near-perfect protection against closed-frame edits and strong resistance to open-frame recontextualizations. For example, under FLUX.2, it reduced human-judged edit success to a total of just 12 out of 300 samples using a tight perturbation budget ($\epsilon = 4$).
- Robust Transferability: VETO protections optimized on a quantized version of FLUX.2 transferred robustly to the full-precision base model, distilled variants (FLUX.2-Turbo), style LoRAs, and smaller architectural derivatives (FLUX.2-Klein).
Limitations and Future Work
- Transformation Robustness: Like many adversarial cloaks, VETO remains sensitive to certain image corruptions applied before editing. While targeted data augmentation during optimization can mitigate horizontal flips and crops, resisting heavy JPEG compression under tight perturbation budgets remains an open challenge.
- Optimization Overhead: VETO requires per-image optimization, taking approximately 57 seconds per image when hooking the double-stream blocks.
- Multi-Reference Complexity: While preliminary results show VETO can protect an image even when mixed with unprotected reference images, more complex multi-image editing environments warrant deeper study.
The authors conclude that truly effective image protection will ultimately require layered defenses, combining user-side cloaking like VETO with inference- and parameter-level model guardrails.
Summary updated