Empowering Anomaly Detection for Industrial Internet of Things with Intelligent Edge Computing and Decentralized Machine Learning

The Industrial Internet of Things (IIoT) aims to connect machinery, equipment, human resources, and all other relevant components in an industrial environment to achieve op-erational and management efficiency. Alongside the potential benefits of IIoT, this field also has many exploitable vulnerabilities. IIoT devices are often limited in resources, making implementing security solutions directly on these devices difficult. Instead, N et-work Intrusion Detection Systems (NIDS) are typically deployed at edge nodes in IIoT systems. In this study, we build a NIDS on cloud infrastructure to simulate edge devices within the Swarm Learning architecture. In this system, we train two models, CNN and RNN, on the Edge-IIoTset dataset. The results show that the CNN and RNN models in the Swarm Learning architecture achieve good performance compared to Centralized Learning models and Federated Learning approaches. Specifically, our system achieves accuracy close to centralized machine learning models and outperforms Federated Learning approaches in binary and multi-label attack detection. Additionally, we evaluate the aggregators in Swarm Learning. The mean aggregator is stable, while the coord and geo aggregators achieve high accuracy in classes with few data points.