Evaluating Cloud Assisted Machine Learning for Internet-of-Things Edge Device Security

Security solutions based on machine learning enabled Intrusion detection systems (ML-IDS) have been popular for IoT edge device security. The cloud support has been considered to leverage heavy workloads off IoT edge devices such as learning, model fitting and updating. In this paper, we apply ML model optimization techniques as 3 distinct key roles of the cloud layer in Machine Learning as a Service (MaaS) architecture. In order to influence the performance of IoT edge devices, the optimization would require: a) device benchmarking that quantifies the available computing resources and network traffic for the edge device; b) feature weighing that determines the contribution of each extracted feature to the overall ML-IDS operation; and c) dimensionality reduction of important features in the training dataset. We examine feature importance and dimensionality reduction techniques on 3 public IDS datasets and present how they affect ML model performance.