Network Coding for Federated Learning Systems

Lingwei Kong, Hengtao Tao, Jianzong Wang, Zhangcheng Huang, Jing Xiao

Abstract: Nowadays, artificial intelligence is limited by privacy and security problems. Compared with the ordinary machine learning, federated learning (FL) enables multiple participants to collaboratively learn a shared machine learning model while keeping all the training data on local devices. However, most of the current secured federated learning systems (FLSs) are built up with high computational and communication costs. On the other hand, optimizing the network structure of federated learning systems can reduce communication complexity by considering the correlation of the transmission channels. In this paper, we propose Network Coding Federated Learning Systems (NC-FLSs). Specifically, it considers the whole communication network by connecting all the clients and the server. Applying a linear NC scheme to construct a linear combination of the original messages, which is transmitted over the network instead of the messages themselves. Based on NC-FLSs, the communication cost is halved and both data privacy and security are improved with the imperceptibly higher computational cost. Moreover, considering that the network coding structure is independent of the FL model, any FLSs can also be upgraded to its corresponding NC-FLSs. We also implement differential privacy on an NC-FLS to train an image classifier while keeping clients’ local data secure and private, which achieves superior performance and efficiency.