On effectiveness of AI-based misbehavior detection in medical IoT

Abstract

Artificial Intelligence (AI) classification techniques are pivotal for misbehavior detection in the Internet of Things (IoT), but their potential for severe failure poses a risk in safety-critical applications. This work introduces a novel statistical methodology to evaluate the operational readiness of these AI systems by quantitatively forecasting their effectiveness throughout the learning process. The significance of our methodology lies in its ability to provide predictive insights into an AI detector’s performance, enabling a system architect to make data-driven decisions about deployment. We use two lightweight statistical analysis methods: one to model device compliance and forecast the detector’s false negative probability ($p_{fn}$) of missing a malicious device and its false positive probability ($p_{fp}$) of misidentifying a benign one, and another to model the learning curve and predict the future misclassification rate. This framework allows a designer to determine precisely when a system has been trained sufficiently to meet predefined safety and reliability targets. We demonstrate the feasibility of our approach on an artificial pancreas system with a smart Continuous Subcutaneous Insulin Infusion (CSII) device, confirming the effective and predictable detection of sophisticated attacks.