P8_22: Feedback Control of Hardware Security Countermeasures with Deep Reinforcement Learning
Topic Areas: Formal Verification / Methods, Integrity Monitoring, Risk Mitigation
Principal investigator: Dr. Afsaneh Doryab, University of Virginia
Co-Principal investigator(s): Dr. James H. Lambert, Virginia
PI Email
Abstract:
Various CHEST technologies involve human experts to specify the properties that must be satisfied for an on-device program to be considered secure. While progress is being made to verify these properties, methods are needed to assist in writing the specifications to be verified. In particular, new technologies could aid human experts to improve the accuracy and completeness of specifications for security countermeasures. This project will develop an interactive system in which a Deep-Reinforcement-Learning (DRL) agent and human expert collaborate on generating, refining, predicting, and assessing the states of security of hardware. For real-world DRL systems, a human expert must account for several layers of complexity. We introduce an approach that identifies complex countermeasure specifications that account for the relationships between features. After these specifications are learned, the interactive DRL tool presents them to a human expert in an interpretable form so the specifications can be encoded and used with any DRL method that guarantees security. The IAB-sourced testbeds will be selected from electric-vehicle charging, battery management, power distribution, navigation, medical, and autonomous systems.