Reproduce “Boosting Gradient Leakage Attacks: Data Reconstruction in Realistic FL Settings” 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
Mingyuan Fan, Fuyi Wang, Cen Chen, Jianying Zhou
FEDLEAK, an enhanced gradient leakage attack, reconstructs client training data with high fidelity even under realistic federated learning conditions like larger batch sizes and complex models. The method achieves this by addressing optimization challenges through partial gradient matching and gradient regularization, revealing that current FL deployments are more vulnerable than previously assumed, particularly for sensitive data like medical images.
Key insights
- The gradient matching problem for data reconstruction is often over-constrained, implying a unique solution, contrary to previous assumptions of multiple solutions.
- Previous gradient leakage attacks failed due to optimization difficulties stemming from an unstable convergence ratio (2μ/L) in the loss landscape, rather than inherent ambiguity.
- Selecting and optimizing only a subset of high-magnitude gradient elements significantly improves the optimization landscape and convergence for data reconstruction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.