Byzantine-Robust FedDG

Byzantine-Robust Federated Domain Generalization

Click here for the Project Report.

  • Designed a scenario in federated learning where both o.o.d data and poisoned data exist in the clients; Showed the SOTA byzantine-robust federated learning algorithm will fail to generalize in such scenario by being too conservative.
  • Proposed a novel federated learning algorithm that encourages the features to align across the domains; the proposed method makes trade-off between domain generalization and byzantine robustness.