In-Network Aggregation for Privacy-Preserving Federated Learning

Abstract

Cross-silo federated learning becomes popular in various fields due to its great promises in protecting training data. By carefully examining the interaction among distributed training nodes, we find that existing federated learning still suffers from security weakness and network bottleneck during model synchronization. It has no protection on training models, which also contain significant private information. In addition, many evidences have shown that model synchronization over wide-area network is slow, bottlenecking the whole learning process. To fill this research gap, we propose a novel cross-silo federated learning architecture that can protect both training data and model by using homomorphic encryption (HE). Instead of sharing the model parameters in plaintexts, we encrypt them using the HE, so that they can be aggregated in ciphertexts. In order to handle the inflated network traffic incurred by HE, we apply the in-network aggregation by exploiting the strong capability of programmable switches. A fast algorithm that jointly considers in-network aggregator placement and traffic engineering has been proposed and evaluated by extensive simulations.