Two-Stage Aggregation based Federated Learning (TSA-FL) for Industrial Internet of Things

Federated learning is a privacy-focused machine learning technique that promises to improve the Industrial Internet of Things (IIoT) systems. It can be used to enhance predictive maintenance, quality control, energy management, supply chain optimization, and anomaly detection. However, the implementation of federated learning in IIoT faces several challenges, including security while sharing parameters, communication overhead, data heterogeneity, and edge device resource limitation. This paper proposed a Two-Stage Aggregation Federated Learning (TSA-FL) technique, where parameter aggregation is carried out in two phases to detect malicious nodes and minimize delays. During a specific round, parameters intercepted by the attacker and poisoned by Byzantine nodes may arrive later than legitimate ones, or they might not be received at all. Hence, the local server sets two time slots namely time-slot-1 for faster arriving first group of edge devices and time-slot-2 for late arriving second group of edge devices. Any device arriving after the time-slot-2 won’t be accepted and considered malicious. The initial time slots are estimated manually, but they can be adjusted later based on the available bandwidth and resources of edge devices. In some cases, Byzantine nodes and attackers with better computational resource may still be able to submit their manipulated parameters within the designated time slots (either time-slot-1 or time-slot-2). This results in another vulnerability in the aggregation process. In this scenario, we propose an anomaly detection mechanism for distinguishing abnormal parameters. Experimental evaluation of the proposed approach shows significant performance improvement by maintaining processing time over the existing state-of-the-art.