Heterogeneous Semi-Asynchronous Federated Learning in Internet of Things: A Multi-Armed Bandit Approach

Federated learning (FL) has recently received significant attention in Internet of Things, due to its capability of enabling multiple clients to collaboratively train machine learning models using neural networks, without sharing their privacy-sensitive data. However, due to the heterogeneity of clients in their computation and communication capability, they might not return the training model to the server at the same time, which may result in high waiting latency at the server, especially in synchronous FL. Although asynchronous FL can reduce the waiting latency, aggregating global model in a completely asynchronous way may lead to some local models out of date, resulting in low training accuracy. To address the above issues, this paper aims to propose a novel Heterogeneous Semi-Asynchronous FL mechanism, named HSA_FL . Firstly, we use a Multi-Armed Bandit (MAB) approach to identify the heterogenous communication and computation capabilities of clients, based on which, we assign different training intensities to clients. Generally, the clients with lower capabilities will be assigned with less number of local updates. In addition, instead of waiting all the clients to return their training models or immediately aggregation after getting a single local model, this paper proposes two aggregation rules, named adaptive update and fixed adaptive, respectively. Finally, simulation results show that the proposed scheme can effectively reduce the training time and improve the training accuracy as compared with some benchmark algorithms.