ACSFL: An adaptive client selection-based Federated Learning with personalized differential privacy for heterogeneous AIoT environments

Abstract

Driven by the rapid development of Artificial Intelligence (AI) and the Internet of Things (IoT), the Artificial Intelligence of Things (AIoT) is increasingly applied in smart environments. Federated Learning (FL) meets the need for intelligent data processing in these environments by providing powerful training capabilities while preserving privacy. However, AIoT environments pose new challenges for FL, particularly due to the heterogeneity of edge devices, which vary in hardware, software, network conditions, and data distribution. These factors degrade model performance and hinder convergence. Additionally, communication overhead and data privacy risks are also critical concerns. Although Differential Privacy (DP) can offer protection, they often apply uniform privacy levels, overlooking the diversity of AIoT devices. On the other hand, while current client-selection approaches partially address the heterogeneity of AIoT devices, they also tend to ignore the impact of the noising mechanisms. In this paper, we propose ACSFL, an adaptive client selection-based FL framework that integrates personalized local DP. By a novel, dynamic evaluation metric of node heterogeneity, privacy budget, and contribution, ACSFL can jointly optimize model performance, privacy preservation, and communication efficiency. We further propose a personalized local differential privacy mechanism in ACSFL, to filter and allocate each client’s budget per round, substantially enhancing privacy preservation and yielding significant accuracy gains under identical overall privacy constraints. All the above assertions are also well supported by theoretical and experimental demonstration. Specifically, our experiments show that ACSFL improves model convergence and generalization by 14% on average, achieves comparable model accuracy with 20% fewer clients, reduces communication overhead by over 25%, and saves about 26% of the privacy budget compared to other client selection methods.