Lightweight Federated Learning for Terminal-Edge-Cloud with Two-Step Privacy Protection in Consumer Electronics

Chen, Chao and Xu, Xiaolong and Zhang, Yifan and Bilal, Muhammad and Xiang, Haolong (2025) Lightweight Federated Learning for Terminal-Edge-Cloud with Two-Step Privacy Protection in Consumer Electronics. IEEE Transactions on Consumer Electronics. ISSN 0098-3063

Text (Lightweight_Federated_Learning_for (1))
Lightweight_Federated_Learning_for_1_.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (1MB)

Abstract

With the rapid development of Internet of Things (IoT) technology, consumer electronics increasingly integrate edge computing to improve real-time decision-making in applications such as smart cities and intelligent transportation. The protection of user privacy has become a critical concern due to the vast amounts of sensitive data generated and processed in consumer electronics. Existing privacy-preserving methods primarily address two aspects: safeguarding sensitive user data during offloading from devices to edge servers, and protecting model parameters in cloud processing. However, these approaches often fail to comprehensively address privacy risks in terminal-edge-cloud collaboration, including challenges arising from data heterogeneity, risks of user privacy leakage, and parameter reverse inference attacks. To address these challenges, we propose a lightweight privacy-preserving federated learning algorithm with two-step differential privacy for consumer electronics, integrated into a terminal-edge-cloud architecture (TECDP). TECDP utilizes edge computing and differential privacy to reduce data transmission and perform local preprocessing and encryption, balancing privacy protection with data utility. Lightweight CNN models run on edge devices, while more complex models are deployed in the cloud for improved accuracy. We conducted extensive experiments on the MNIST and CIFAR datasets and evaluated the impact of varying privacy budgets and parameters on model performance. The results demonstrate that TECDP maintains high accuracy while reducing the risk of data leakage.