Issues and Challenges in Machine Learning-based IoT Device Discovery: An Empirical Study

IoT device discovery and identification is a requisite part of ensuring strong cyber security posture for homes, organizations & industries. Increased use of encryption by such devices poses a present and future challenge for traditional identification methods based on Deep Packet Inspection (DPI). Recent research has proposed the use of AIbased approaches as a solution. However, enabling such systems to operate accurately and scalably across heterogeneous IoT devices and networks remains a challenge. This paper makes multiple contributions to advance the state of the art. First, we analyze representative domain research and summarize results. Second, we discuss pending issues and challenges which emerge from existing research. Third, we experimentally investigate and validate the pending challenges using three public datasets as well as two private datasets with 30 IoT devices. Our findings indicate that while research around per-device IoT ML models has advanced significantly, further research is required in the area of generalized AI models which carry out accurate discovery & identification across different $I o T$ devices and networks. We identify in particular, the area of device category identification as deserving of further research attention.