Secure Gradient Aggregation for Wireless Multi-Server Federated Learning.

In this paper, we investigate the secure gradient aggregation problem for K-server wireless Federated learning (FL) systems. We propose a coded aggregation scheme such that any set of up to T colluding servers cannot infer any information about the local updates, including the aggregation value. In our scheme, each user encodes its local update to K confidential messages using Lagrange Coding, and then sends one confidential message to each server using an artificial noise alignment approach. In the downlink, each server delivers the summation of the confidential messages, by which every user can recover the aggregation of local updates. For the proposed scheme, we characterize the uplink and downlink communication latency, and show that the communication latency monotonically decreases with the total number of servers K while increasing with the number of colluding servers T.