An Enhanced Privacy-Preserving and Poisoning-Resilient Federated Learning Scheme for Heterogeneous Intelligent Internet of Things

Abstract

Machine learning and privacy preservation are key technologies driving the development of Intelligent Internet of Things (IIoT). Federated learning (FL) has gained widespread attention for enabling data privacy while supporting collaborative learning in IIoT systems. However, real-world FL deployments are still vulnerable to critical security threats and performance bottlenecks, including model inversion attacks, model poisoning attacks, and device heterogeneity issues. To address these challenges, we propose an enhanced privacy-preserving federated learning scheme (EPFL) designed for IIoT scenarios. It is designed to maintain privacy and robustness of the FL system under extreme client heterogeneity. The CKKS homomorphic encryption scheme is employed to secure gradient information. An innovative monitoring server is introduced, where encrypted gradients are evaluated through a three-party Shamir secret sharing protocol. This design preserves the confidentiality of sensitive data even when one party is compromised. A multi-metric client scoring and grouping framework is employed by the monitoring server. Manhattan distance, cosine similarity, and Jaccard similarity are integrated to dynamically evaluate client contributions in device heterogeneous environments. This scheme enables the detection of malicious or low-performing clients (stragglers) while ensuring both the accuracy and rapid convergence of the global model. Experiments show that EPFL limits accuracy loss under 40% malicious clients to below 2%, removes over 95% of stragglers within 20–50 rounds, and reduces straggler inclusion by about $1.18{\times }$ compared to random selection, while maintaining high model accuracy.