Reproduce “Data Leakage in Federated Averaging” 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
Dimitar I. Dimitrov, Mislav Balunović, Nikola Konstantinov, Martin Vechev
Researchers from ETH Zurich developed the first successful data leakage attack against Federated Averaging (FedAvg), the most widely used federated learning algorithm, demonstrating that client data can be reconstructed from aggregated model updates. The attack achieved up to 65.5% image reconstruction success on FEMNIST with 21.5 PSNR and 58.8% on CIFAR100 with 19.8 PSNR in realistic settings.
Key insights
- Federated Averaging is vulnerable to data leakage, dispelling the notion that its architectural complexities inherently provide strong privacy guarantees.
- Simulating the full local training process on the server side, including intermediate weight updates, is crucial for effectively reconstructing client data from FedAvg updates.
- Leveraging the invariance of client data across multiple local epochs through an order-invariant prior significantly improves the quality and robustness of reconstructed images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.