Reproduce “ZIUM: Zero-Shot Intent-Aware Adversarial Attack on Unlearned Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hyun Jun Yook, Ga San Jhun, Jae Hyun Cho, Min Jeon, Donghyun Kim, Tae Hyung Kim, Youn Kyu Lee
ZIUM introduces a zero-shot intent-aware adversarial attack method that induces unlearned text-to-image diffusion models to generate forbidden concepts, allowing for user-specified customization. The method achieved significantly higher attack success rates, up to 62.0 percentage points higher than existing approaches, and reduced attack time by up to 149.5 times through its zero-shot capability.
Key insights
- Explicit visual content from a target image is more effective in inducing unlearned concepts in generative models than text-only prompts.
- Optimizing a visual-text alignment module enables a zero-shot adversarial capability, allowing for highly efficient, customized attacks against the same unlearned concept.
- Even after machine unlearning, generative models remain vulnerable to subtle adversarial manipulations that combine unlearned concepts with user-specified contexts or styles.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.