The rapid advancement of the Internet of Things (IoT) has transformed edge devices from simple data collectors into intelligent units capable of local processing and collaborative learning. However, the vast amounts of sensitive data generated by these devices face severe constraints from “data silos” and risks of privacy breaches. Federated learning (FL), as a distributed collaborative paradigm that avoids sharing raw data, holds great promise in the IoT domain. Nevertheless, it remains vulnerable to gradient leakage threats. While traditional differential privacy (DP) techniques mitigate privacy risks, they often come at the cost of significantly reduced model performance—a limitation particularly pronounced in resource-constrained IoT environments characterised by non-independent and identically distributed (non-IID) data distribution. To bridge the gap between privacy preservation and high performance on heterogeneous data, this paper proposes a novel personalized federated learning (PFL) method, FedGLP-ADP. This method leverages historical gradient information to provide a more detailed partitioning of parameters, aiming to prevent personalized knowledge from being affected by noise as much as possible, thereby reducing model degradation. Building on this, we propose an adaptive DP mechanism that optimizes both the clipping and noising steps to minimize the impact of noise on global knowledge. Experimental results show that FedGLP-ADP exhibits superior performance compared to other representative methods under different privacy levels and non-IID degrees.
Xiao et al. (Thu,) studied this question.