DSAFL:Decentralized secure aggregation with communication path optimization for cross-silo federated learning

Abstract

Cross-Silo Federated Learning (CSFL) facilitates collaborative machine learning (ML) across organizations by locally training models and centrally aggregating model updates. Currently, this approach is shifting to decentralized aggregation due to the limitations of centralized aggregation such as single-point failures and network congestion. However, existing decentralized aggregation methods often suffer from privacy leakage and high communication cost. To address these issues, we propose DSAFL, a decentralized secure aggregation scheme for CSFL. In DSAFL, we present a staged secure aggregation method based on multi-key homomorphic encryption, which enables load-balanced collaborative aggregation computation across clients while preserving model update confidentiality and providing verifiability of the aggregation result. DSAFL optimizes communication paths by jointly considering communication cost and reliability, enabling cost-efficient and robust secure aggregation across diverse network topologies, and further reduces communication cost through non-interactive decryption. The security analysis proves that DSAFL is semi-honestly secure and resistant to client collusion attacks. The experimental results confirm the practicality and applicability of DSAFL, and show significant advantages in both accuracy and privacy. With a combination of computational balancing, low communication cost, and privacy preservation, DSAFL provides a solution for enabling sustainable ML collaboration across organizations.