Towards Trustworthy IIoT: A Reinforcement Learning Framework with Federated Unlearning

The Industrial Internet of Things (IIoT) enhances automation and efficiency but faces challenges from data poisoning, device failures, and unstable networks. To address these issues, we propose a unified framework, FU-MARL, that integrates federated unlearning with multi-agent reinforcement learning (MARL) to improve system resilience and adaptability. The framework comprises intelligent agents, a scheduling server, and a secure communication network. Each agent performs local model training, threat detection, and unlearning through a federated learning module and a reinforcement learning module for adaptive control. The server coordinates global aggregation, verifies unlearning results, and ensures integrity, while encrypted communication safeguards data privacy. The unlearning process can be triggered by opt-out requests, attacks, or administrative commands, enabling the system to remove targeted data contributions without full retraining. Results show that the proposed approach effectively reduces retraining costs and enhances robustness in dynamic IIoT environments.