Using Reinforcement Learning to Spoof a Monitored Kalman Filter
Abstract
Modern hardware systems rely on state estimators such as Kalman filters to monitor key variables for feedback and performance monitoring. The performance of the hardware system can be monitored using a chi-squared fault detection test. Previous work has shown that Kalman filters are susceptible to false data injection attacks. In a false data injection attack, intentional noise and/or bias is added to sensor measurement data to mislead a Kalman filter in a way that goes undetected by the chi-squared test. This thesis proposes a method to deceive a Kalman filter where the attack data is generated using reinforcement learning. It is shown that reinforcement learning can be used to train an agent to manipulate the output of a Kalman filter via false data injection and without being detected by the chi-squared test. This result shows that machine learning can be used to successfully perform a cyber-physical attack by an actor who does not need to have in-depth knowledge and understanding of mathematics governing the operation of the target system. This result has significant real-world impact as modern smart power grids, aircraft, car, and spacecraft control systems are all cyber-physical systems that rely on trustworthy sensor data to function safely and reliably. A machine learning derived false data injection attack against any of these systems could lead to an undetected and potentially catastrophic failure.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2022
- Accession Number
- AD1184770
Entities
People
- Dylan A. Bonitz
Organizations
- Naval Postgraduate School