Towards value-sensitive and poisoning-proof model aggregation for federated learning on heterogeneous data

Federated Learning (FL) enables collaborative model training without sharing data, but traditional static averaging of local updates leads to poor performance on heterogeneous data. The following remedies, either by scheduling data distribution or mitigating local discrepancies, predominately fail to handle fine-grained heterogeneity (e.g., local imbalanced labels). To commence, we reveal that static averaging leads to the global model suffering from the mean fallacy. That is, the averaging process favors the local model with large parameters numerically rather than knowledge. To tackle this, we introduce FedVSA, a simple-yet-effective model aggregation framework sensitive to heterogeneous local data merits. Specifically, we invent a new global loss function for FL by prioritizing the valuable local updates, facilitating efficient convergence. We deduce a softmax-based aggregation rule and prove its convergence property via rigorous theoretical analysis. Additionally, we expose poisoning threats of model replacement that utilize the mean fallacy for attacks. To mitigate this threat, we propose a two-step mechanism involving auditing historic local training statistics and analyzing the Shapley Value. Through extensive experiments, we show that FedVSA achieves faster convergence (~1.52×) and higher accuracy (~1.6%) compared to the baselines. It also effectively mitigates poisoning attacks by agilely recovering and returning to normal aggregation.