Security Analysis of RL-Based Artificial Pancreas Systems

Reinforcement learning (RL) based models promise to replace time-consuming traditional model-based control methods for medical control systems. Recently, Deep RL approaches have been explored for use in autonomous systems for glycemic control, often termed Artificial Pancreas systems, which require closed-loop communication between a glucose sensor and an insulin pump. In this work, we investigate the robustness of RL4BG, a prominent deep reinforcement learning-based AP controller, to a suite of glucose sensor malfunctions. We model two classes of realistic malfunctions stemming from natural and/or adversarial factors: a Denial-of-Service failure class that simulates a worst-case sensor malfunction, and a Subtle manipulations failure class that simulates stealthier but prolonged failure. Our findings show that this new class of medical control systems may be vulnerable to anomalous inputs in safety-critical settings. These vulnerabilities motivate further work into training medical RL systems in an adversarially robust fashion.