LIGHTWEIGHT AI MODELS FOR IDSIN IIOT : BALANCING ACCURACY ANDCOMPUTATIONAL EFFICIENCY

The Industrial Internet of Things (IIoT) is quickly changing industries like manufacturing, transportation,and energy. It allows real-time monitoring, automation, and predictive analytics, makingprocesses more efficient. But with more connectivity comes greater cybersecurity risks. Many IIoTdevices have limited hardware capabilities, making them easy targets for cyberattacks. TraditionalIntrusion Detection Systems (IDS) [1] are often unsuitable for such environments due to their highcomputational demands.This Thesis investigates the use of lightweight machine learning (ML) models for IDS, specificallydesigned for IIoT systems. The purpose is to explore how ML techniques can be adapted torun efficiently on constrained devices without sacrificing detection performance in terms of accuracy.The main research questions focus on balancing accuracy and computational efficiency andunderstanding the potential of TinyML-based approaches 1 for IIoT security.A simulation-based methodology is adopted using the TON_IoT dataset 2, a rich and heterogeneousdataset specifically designed to reflect real-world IIoT scenarios. Models including a ConvolutionalNeural Network (CNN) and a Random Forest classifier were implemented, preprocessedusing standard machine learning pipelines, and optimized using techniques such as quantizationand pruning.The simulation results show that lightweight AI models can achieve high accuracy while maintaininglow computational overhead, making them feasible for embedded deployment. This studycontributes to the growing field of IIoT security by offering a practical and efficient approach tobuilding IDS for critical infrastructure environments. The work also highlights the importance ofwell-structured datasets and optimization techniques for future research in this area.