BEV-SGD: Best Effort Voting SGD Against Byzantine Attacks for Analog-Aggregation-Based Federated Learning Over the Air

As a promising distributed learning technology, analog aggregation based federated over the air (FLOA) provides high communication efficiency and privacy provisioning under edge computing paradigm. When all devices (workers) simultaneously upload their local updates to parameter server (PS) through commonly shared time-frequency resources, PS obtains averaged update only rather than individual ones. While such concurrent transmission scheme reduces latency costs, it unfortunately renders FLOA vulnerable Byzantine attacks. Aiming at Byzantine-resilient FLOA, this paper starts from analyzing channel inversion (CI) mechanism that is widely used for power control in FLOA. Our theoretical analysis indicates although CI can achieve good performance benign scenarios, fails work well with limited defensive capability against Then, we propose novel called best effort voting (BEV) policy integrated stochastic gradient descent (SGD). BEV-SGD enhances robustness of attacks, by allowing workers send maximum transmit power. Under worst-case derive expected convergence rates BEV policies, respectively. The rate comparison reveals our outperforms its counterpart terms better behavior, which verified experimental simulations.