FLANDERS: A Byzantine-Resilient Aggregation Scheme for Federated Learning via Matrix Autore- gression on Client Updates

The thesis focuses on a specific type of attack called Byzantine, where a subset of malicious clients participating in the FL network attempts to poison the global model by sending arbitrarily corrupted local updatesto the server. Without a robust aggregation scheme at the server's end, the global model may get significantly jeopardized or even fail to converge. Thus, we propose FLANDERS, a novel FL aggregation rule robust to Byzantine attacks. FLANDERS considers the model updates sent by clients at each FL round as a matrix-valued time series. Then, it identifies malicious clients as outliers of this time series by comparing actual observations with those estimated by a matrix autoregressive forecasting model. Experiments conducted on several datasets demonstrate that FLANDERS remains highly effective even under extremely severe attack scenarios, as opposed to existing defense strategies.