Reproduce “Upcycling Noise for Federated 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
Jianan Chen, Qin Hu, Fangtian Zhong, Yan Zhuang, Minghui Xu
Researchers at [Authors' institutions not provided in report] introduce Federated Unlearning with Indistinguishability (FUI), a scheme designed for Differential Privacy-enhanced Federated Learning (DPFL) that leverages existing DP noise for efficient unlearning. The method achieves statistical indistinguishability for forgotten data, demonstrating superior efficiency and comparable model utility to adapted state-of-the-art unlearning approaches.
Key insights
- The noise inherent in Differential Privacy (DP) can be strategically leveraged, or 'upcycled', to actively contribute to the unlearning process, rather than solely being an impediment.
- Quantifying unlearning effectiveness through 'epsilon-indistinguishability' provides a rigorous, measurable guarantee that the unlearned model does not retain discernible information beyond a specified privacy budget.
- A game-theoretic approach can effectively resolve conflicts of interest between the server and clients, providing optimal and consistent strategies for unlearning parameters in DPFL.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.