BioMutFed+: Mutation-Driven Federated Learning for IIoT

Abstract

The evolution towards Industry 5.0 underscores the critical need for privacy-preserving and human-centric technologies within Industrial Internet of Things (IIoT) ecosystems. Federated Learning (FL), an essential framework enabling decentralized model training while safeguarding data privacy, continues to encounter significant obstacles such as non-IID data distributions, adversarial threats, and limited scalability. To address these issues, we propose <sc><bold>BioMutFed+</bold></sc>, a biologically inspired federated learning framework featuring adaptive mutation-based gradient perturbations, pheromone-driven adaptive client selection, and robust trimmed-mean aggregation. Comprehensive theoretical analysis guarantees convergence at a rate of <inline-formula><tex-math notation="LaTeX">$\mathcal {O}(1/\sqrt{T})$</tex-math><alternatives><mml:math><mml:mrow><mml:mi mathvariant="script">O</mml:mi><mml:mo>(</mml:mo><mml:mn>1</mml:mn><mml:mo>/</mml:mo><mml:msqrt><mml:mi>T</mml:mi></mml:msqrt><mml:mo>)</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="hawbani-ieq1-3707971.gif"/></alternatives></inline-formula>. Empirical evaluations on Digits (visual tasks), NASA FD004 (predictive maintenance), and Wisconsin Breast Cancer datasets demonstrate rapid convergence within 20 rounds, variance reductions of 1.2–2.9× against established FL methods, and enhanced adversarial robustness under 20% malicious client scenarios compared to current approaches. Furthermore, we introduce <sc>FedMutAdam</sc>, a lightweight <sc>BioMutFed+</sc> variant integrating adaptive gradient clipping with server side Adam style updates, optimized for resource-constrained IIoT devices. Our framework thus represents a scalable and robust solution advancing secure and adaptive intelligent systems for IIoT systems.