Reproduce “Robbing the Fed: Directly Obtaining Private Data in Federated Learning with Modified Models” 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

Liam Fowl, Jonas Geiping, Wojtek Czaja, Micah Goldblum, Tom Goldstein

Researchers from the University of Maryland and NYU developed a method for a malicious federated learning server to directly recover verbatim user data by subtly modifying the model architecture with an "imprint module." This technique allows high-fidelity reconstruction of individual data points even from large aggregated batches, achieving average PSNRs of 75.75 for batches of 64 images and 161.36 for single image recovery from batches of 16,384 images, and also works effectively for text data.

Key insights

  • A malicious server, by subtly controlling the model architecture, can directly and analytically extract verbatim user data from federated learning gradient updates, even in large aggregated batches.
  • The "imprint module" strategy circumvents the limitations of prior gradient inversion attacks, which fail for large batch sizes, by uniquely partitioning the data space.
  • The attack is data-agnostic, demonstrating efficacy across diverse data modalities like images and text, without requiring domain-specific priors for reconstruction.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.