Autonomous Network Intrusion Detection for Resource-Constrained Devices of the Internet of Things

With the Internet of Things (IoT), cars, home assistants, cameras and other devices may be part of an ubiquitous network with access to personal information in real-time, able to interact with the environment and even influence people. For this reason, empowering such devices with network intrusion detection algorithms might be vital for the development of a more secure and trustworthy Internet of Things. In this work we study the performance and accuracy of machine learning models trained for this task, and analyzed the impact of their deployment in a microcontroller-based System-on-Chip used by IoT devices, along with the impact of TinyML-based optimizations, such as vectorization and quantization. From our experiments, Decision Trees presented very low inference time of 5 microseconds on average, and higher accuracy of 99.9% when compared to Logistic Regression and Neural Networks, being a viable solution for real-time, accurate and autonomous Network Intrusion Detection system for IoT devices.