Aldp-fl: an adaptive local differential privacy-based federated learning mechanism for IoT

Federated learning offers an effective solution for safeguarding data privacy in the Internet of Things ecosystem among diverse stakeholders. To enhance data usability for sharing and collaboration, the integration of differential privacy (DP) techniques into federated learning becomes crucial, providing essential support for system sustainability. In fact, differential privacy-based federated learning (DPFL) has gained widespread application across various domains, including healthcare, finance, and smart homes. However, traditional DPFL faces challenges, such as potential privacy leakage due to the plaintext transmission of intermediate content between the central server and clients, as well as the adverse impact of DP on model accuracy. In this article, we propose ALDP, a federated learning-based adaptive local differential privacy mechanism for IoT, which aims to address the privacy leakage and model accuracy degradation problems encountered by traditional DPFL. Specifically, we employ a SAM optimizer to mitigate the negative impact of DP through perceptual tuning and gradient normalization. We further implement an effective threshold cropping technique to manage gradient explosion and sparsity, and apply hierarchical adaptive noise to ensure balanced privacy protection across both data and participants. Experimental results demonstrate that our scheme maintains high model accuracy while preserving privacy. Compared to existing methods, our ALDP mechanism achieves a training and testing accuracy difference of only 8.47% on the EMNIST dataset, significantly outperforming other benchmark methods.