Training Aggregation in Federated Learning

Abstract

Federated learning is a new machine learning paradigm for distributed data. It enables multi-party cooperation to train global models without sharing their private data. In the classic federated learning protocol, the model parameters are the interaction information between the client and the server. The client can update the local model according to the global model parameters, and the server can aggregate the updated model parameters of each client to obtain a new aggregation model. However, in the actual federated learning scenario, there are still privacy problems caused by model stealing attack in collaborative learning using model parameters as interactive information. Therefore, we use knowledge distillation technology to avoid the model stealing attack in federated learning, and on this basis, we propose a novel aggregation scheme, which can make the output distribution of each customer refine the aggregation results through model training. Experiments show that the scheme can achieve normal convergence while ensuring privacy security, and reduce the number of interactions between client and server, thus reducing the resource consumption of each client participating in federated learning.