Robust Aggregation for Federated Learning

We present a novel approach to federated learning that endows its aggregation process with greater robustness potential poisoning of local data or model parameters participating devices. The proposed approach, Robust Federated Aggregation (RFA), relies on the updates using geometric median, which can be computed efficiently Weiszfeld-type algorithm. RFA is agnostic level corruption and aggregates without revealing each device’s individual contribution. establish convergence robust algorithm for stochastic additive models least squares. also offer two variants RFA: faster one one-step aggregation, another on-device personalization. experimental results deep networks three tasks in computer vision natural language processing. experiments show competitive classical when low, while demonstrating under high corruption.