Reproduce “Model Inversion Attack against 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
Lei Zhou, Youwen Zhu
Lei Zhou and Youwen Zhu introduce the Federated Unlearning Inversion Attack (FUIA), a method demonstrating that existing federated unlearning techniques can inadvertently leak private data by reconstructing forgotten features and inferring forgotten class labels. The attack effectively compromises the privacy guarantees of federated unlearning across various scenarios, establishing a new state-of-the-art in exposing these vulnerabilities.
Key insights
- Federated Unlearning processes, despite their privacy goals, inherently create new vulnerabilities that can be exploited for model inversion attacks.
- The differences in global model parameters before and after federated unlearning contain sufficient information to infer or reconstruct data that was supposedly forgotten.
- Effective defenses against FUIA, such as gradient pruning or perturbation, introduce a direct trade-off with the model's overall utility and performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.