A Privacy-Preserving Distributed Machine Learning Aggregation Scheme

Abstract

Federated Learning enables data to remain on local devices, yet malicious servers can still infer sensitive user information by analyzing client-uploaded model updates, posing significant privacy leakage risks. Existing secure aggregation schemes—such as Differential Privacy, Homomorphic Encryption, and traditional Secret Sharing—struggle to achieve an ideal balance among model accuracy, computational/communication overhead, and adaptability to complex aggregation scenarios. To address this, this paper proposes a multi-server secure aggregation scheme based on additive secret sharing. The scheme introduces multiple non-colluding servers; each client splits its local model update into random secret shares and distributes them to these servers. Each server independently performs weighted aggregation, and the global model is reconstructed by a leader server. Theoretical analysis demonstrates that, owing to the linear homomorphism of additive secret sharing, the aggregation result of our scheme is mathematically equivalent to the standard FedAvg algorithm. Experimental results on the MNIST dataset show that our scheme achieves a test accuracy of 86.13% after 30 global rounds, closely matching the baseline FedAvg (86.12%), with only a 6% additional time overhead per round. Under reasonable non-collusion assumptions, the scheme achieves information-theoretic privacy protection, effectively breaking the trilemma among privacy, accuracy, and efficiency.Keywords:Federated Learning, Additive Secret Sharing,Secure Aggregation; Privacy Preservationords.