Secure Mutual Model Assessment for Robust Federated Learning Under Data Heterogeneity

Federated Learning (FL) enables distributed model training while maintaining data privacy, but it often suffers performance degradation due to data heterogeneity among clients. Traditional aggregation methods, relying on centralized servers, are vulnerable to Byzantine attacks and privacy breaches. To address these challenges, we propose a Secure Mutual Model Assessment (SMMA) scheme. SMMA allows federated clients to securely evaluate each other’s models through privacy-preserving inference without exposing local datasets or model parameters. The resulting assessment scores serve as weights for robust model aggregation. Experimental results present that SMMA significantly improves aggregation robustness and model accuracy under highly heterogeneous data distributions.