Evaluation of Lightweight Machine Learning-Based NIDS Techniques for Industrial IoT

Internet of Things (IoT) devices have revolutionized communication, transportation, healthcare, and many other fields. In particular, the adoption of these devices has propelled the growth of Industry 4.0 to an exponential pace. However, while this vast pool of interconnected devices broadens the opportunities for better business and better lives, it also attracts the attention of cybercriminals. Nevertheless, it has been shown that the resource-constrained nature of these devices inhibits the deployment of traditional security measures. To this end, we investigate how various lightweight Machine Learning-based intrusion detection systems (IDSs) can be implemented on resource-constrained IoT devices. Specifically, we train various decision tree and neural network-based models and implement them on Raspberry Pi and Field-Programmable Gate Array (FPGA) platforms. Furthermore, we evaluate our implementations on the IoT-23 and TON IoT datasets and compare the results in terms of classification performance, throughput and resource consumption. We show that tree-based models surpass the neural network-based models in classification performance and throughput but that hardware acceleration on FPGA can aid in closing the gap in terms of throughput. As such, this work opens the path for the deployment of a real-time distributed IDS on low-cost devices.