Reproduce “FedCARE: Federated Unlearning with Conflict-Aware Projection and Relearning-Resistant Recovery” 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
Yue Li, Mingmin Chu, Xilei Yang, Da Xiao, Ziqi Xu, Wei Shao, Qipeng Song, Hui Li
Researchers from Xidian University, RMIT University, and CSIRO developed FedCARE, a federated unlearning framework that efficiently removes data influence while preserving utility and preventing relearning. The method achieved up to a 63x speedup over retraining, maintained accuracy on retained data comparable to retraining, and reduced backdoor attack success rates to 0% by employing conflict-aware projection and a relearning-resistant recovery strategy.
Key insights
- Gradient ascent, when properly constrained, can be a low-cost primitive for effective unlearning, especially when target data is available at the client.
- Data-free synthesis, via model inversion, can provide informative class-level signals to constrain utility degradation during unlearning without violating privacy by accessing retained data.
- Freezing the feature extractor and filtering server updates are crucial to prevent the re-introduction of forgotten knowledge and improve robustness against attacks during model recovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.