Secure Asynchronous Federated Learning for Edge Computing Devices

Abstract

Federated learning (FL) solves the problem of isolated data islands and improves machine learning models by training on large amounts of data collected from multiple parties. However, the traditional synchronous FL training method will lead to the problem of laggards due to the insufficient computing power of some edge computing devices, which in turn affects the accuracy of training and increases communication overhead. Asynchronous federated learning solves the training problem of traditional FL very well. However, traditional privacy-preserving methods for FL are not suitable for asynchronous federated learning, existing asynchronous training schemes cannot guarantee the privacy of data. Faced with this challenge, this paper proposes a weighted secure asynchronous federated learning scheme, WASecAgg, which can ensure training accuracy and data privacy on the premise of guaranteeing the training efficiency of federated learning. The final experiments show that WASecAgg performs better in terms of training efficiency, training accuracy and privacy-preserving of federated learning.