跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2607.27292 · v2

VETO: Towards Protecting Images From Frontier AI Editing

首次提交 2026-07-29 23:03
版本更新 2026-11-19 00:12
雷达观测 2026-11-19 15:04
arXiv ↗PDF ↗

摘要

Modern image editors blur the boundary between image editing and text-to-image synthesis, extracting people and objects from a single photo and inserting them into new scenes without subject-specific fine-tuning. Existing defenses struggle to disrupt such manipulations in these models without large, visibly damaging perturbations to the images. We introduce VETO, a white-box image-cloaking method for open-weight editors that maximizes the entropy of attention between reference image and generation tokens to disrupt selective retrieval of source details, while requiring neither the future editing instruction nor modifications to the model. We also introduce VetoBench, a benchmark of 300 benign and harmful tasks balanced between within-scene editing and recontextualization. Full Pareto evaluations on FLUX.2 and FIBO-Edit across three benchmarks and five perturbation budgets demonstrate overall better protection-fidelity trade-offs, particularly against recontextualization. Protected images remain effective on fine-tuned and distilled variants of the editor without reoptimization.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-19 00:12 · 最新观测版本

VETO: Towards Protecting Images From Frontier AI Editing

Modern image editors blur the boundary between image editing and text-to-image synthesis, extracting people and objects from a single photo and inserting them into new scenes without subject-specific fine-tuning. Existing defenses struggle to disrupt such manipulations in these models without large, visibly damaging perturbations to the images. We introduce VETO, a white-box image-cloaking method for open-weight editors that maximizes the entropy of attention between reference image and generation tokens to disrupt selective retrieval of source details, while requiring neither the future editing instruction nor modifications to the model. We also introduce VetoBench, a benchmark of 300 benign and harmful tasks balanced between within-scene editing and recontextualization. Full Pareto evaluations on FLUX.2 and FIBO-Edit across three benchmarks and five perturbation budgets demonstrate overall better protection-fidelity trade-offs, particularly against recontextualization. Protected images remain effective on fine-tuned and distilled variants of the editor without reoptimization.

查看此版本解读 →查看此版本原文 ↗