Reproduce “CRFU: Compressive Representation Forgetting Against Privacy Leakage on Machine Unlearning” 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

Weiqi Wang, Chenhan Zhang, Zhiyi Tian, Shushu Liu, Shui Yu

CRFU, developed by Wang et al., provides a machine unlearning technique that reduces privacy leakage during data removal by operating on compressed internal representations. The method achieves approximately 200% better defense against reconstruction attacks on MNIST compared to baseline methods while preserving model utility with only 1.5% accuracy degradation.

Key insights

  • Unlearning updates themselves can inadvertently leak private information about the data being removed, presenting a privacy risk.
  • Stronger unlearning, which requires more distinctive model updates, can paradoxically lead to greater potential for information leakage.
  • Compressing internal representations using Information Bottleneck theory can effectively limit information flow from erased data, offering a defense against privacy attacks.

Summary by alphaXiv.

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