Robust Privacy-Enhanced Aggregation Scheme for Federated Learning in Industrial Internet of Things

Federated learning (FL) offers a decentralized approach for collaborative model training, making it a promising solution for data-driven optimization in Industrial Internet of Things (IIoT) environments. However, the sensitive nature of industrial data, combined with privacy concerns and security risks, poses significant challenges for FL deployment in IIoT applications. Specifically, the risks of gradient leakage, malicious client attacks, and insufficient privacy protection hinder the wide adoption of FL in IIoT. To address these challenges, this paper proposes a Robust Privacy-Enhanced Aggregation (RPEA) scheme for federated learning in IIoT, which ensures privacy protection for both model parameters and user data. The RPEA scheme is built on a distributed trust model across two servers, which execute gradient aggregation and model updates without revealing any intermediate results by using secret sharing and secure two-party computation. To further enhance security, we incorporate an l2 norm defense mechanism that protects against gradient boosting attack by securely computing the l2 norm of the gradients and filtering out malicious updates, thus ensuring the robust privacy and integrity of IIoT data. In terms of efficiency, RPEA optimizes both the preprocessing and online computing of l2 norm, significantly reducing both computation and communication overhead. Compared to the baseline method, RPEA achieves approximately a 30-fold speedup in l2 norm computation. Experimental results demonstrate that RPEA effectively defends against gradient inversion and boosting attacks, while maintaining high model accuracy, robustness, and privacy protection in federated learning model aggregation within IIoT environments.