AFL: Adaptive Federated Learning Based on Personalized Model and Adaptive Communication

Abstract

Federated learning (FL) has given indications of being effective as a architecture for distributed machine learning, which can ensure the data security for each client and train the global deep learning model. Due to the rapid development of this technology, issues remain to be fully explored. Among them, the robustness of the system, privacy protection and communication efficiency are the key factors affecting quality of FL. Here we propose a new FL model based on personalized model and adaptive communication, called Adaptive Federated Learning (AFL). Our model mainly uses two mechanisms: a) Each client trains the personalized local model according to its own local data set; b) Adaptive chain communication mode is adopted in federation aggregation to reduce the time spent in synchronizing training result. A large number of experiments on two public datasets: MNIST and CIFAR10 show that our model is more accurate by over 5% compared with the FedAvg, and is also faster by over 10% compared with Chain-PPFL, which provides a very important significance in theory and practical production.