Reproduce “Rethinking the Vulnerability of Concept Erasure and a New Method” 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

Alex D. Richardson, Kaicheng Zhang, Lucas Beerens, Dongdong Chen

A study reveals that concept-erased diffusion models remain pervasively vulnerable to adversarial attacks, a characteristic largely inherited from the pre-unlearned model. Researchers introduce RECORD, a tangential-coordinate-descent algorithm, which achieves up to 17.8 times higher concept restoration rates than prior methods, establishing a stronger benchmark for evaluating unlearning defenses.

Key insights

  • Adversarial vulnerabilities in concept-erased diffusion models are pervasive across the prompt embedding space and are largely inherited from the original, pre-unlearned model.
  • Current concept erasure methods primarily achieve local suppression of concepts rather than global eradication, leaving numerous distant adversarial regions unaddressed.
  • Coordinate-descent methods offer a more effective approach for discrete text optimization in concept restoration compared to projection-based gradient optimization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.