RFL-APIA: A Comprehensive Framework for Mitigating Poisoning Attacks and Promoting Model Aggregation in IIoT Federated Learning

With the development of industrial Internet of Things (IIoT), federated learning (FL) is important for protecting sensitive data from various Internet of Things devices (i.e., clients in FL). Despite FL's privacy benefits, attackers (e.g., untrusted clients) can still compromise the performance of the global model through model poisoning attacks. Unfortunately, two key challenges hinder effective detection and impact the performance of the global model in FL: first, accurately identifying malicious models to defend against attacks, and second, efficiently aggregating local models after detecting malicious clients. To address these challenges, we propose an improved FL system based on fuzzy rules, termed RFL-APIA. Compared to the conventional FL system, we have designed two novel components, federated learning generalized depth detection (FedGDD) and Fedsv-Weighted, to enhance performance and mitigate model poisoning attacks. Specifically, FedGDD introduces variance reduction by examining the relationship between local and global model gradients, thereby mitigating interference in nonindependent and identical distributed settings. It further implements an adaptive penalty factor-based scoring system, leveraging variations in local model updates for precise identification and mitigation of attacks. Based on FedGDD's output, the Fedsv-Weighted mechanism dynamically updates the global model's aggregation weights by considering local models' contributions, thus improving model aggregation. Extensive experiments demonstrate that RFL-APIA effectively prevents model poisoning attacks during training, ensuring model security, and guaranteeing a certain level of accuracy and convergence for the global model.