Counter-Autonomy through Adversarial Reinforcement Learning

This Phase II project aims at addressing the USAF National Defense-related mission need in the area of Next Generation Air Dominance, specifically by countering an adversary’s autonomous system through the exploitation of vulnerabilities in its artificial intelligence capabilities. We believe technology development under this effort could contribute to future mission need fulfillment. The mission impact of this project on the DAF and DoD will be realized in both offensive and defensive tactical air combat applications, and will be extendable to many applications involving autonomous artificial intelligence. The main goals of our involvement in this project are: (1) Exploit and defeat a static autonomous tactical autopilot created through AI methods (2) Enhance the robustness of autonomous tactical autopilots to resist adversarial exploitation As AI-enabled fighters become commonplace, we need to develop methods that can detect flaws or vulnerabilities in the opponent’s AI (commonly referred to as the victim AI).  A USAF-developed AI, ARTUµ, became an active team member of a U-2 aircraft during a flight test from Beale AFB in 2020 and the race to train our AI against adversarial agents has since been a top priority. Commercial environments that operate under similar adversarial conditions include autonomous vehicles, gaming platforms and cybersecurity posturing software tools. Thus, we have identified the appropriate intersection between deep learning and adversarial methods from this trade space to create a novel method that has a dual government and commercial use. The work proposed in this project is to modify the Mobius Logic commercially used automated Machine Learning Operations tools to enable the US Air Force to build a counter-autonomy adversarial reinforcement learning agent. When complete this agent will be utilized in several USAF programs such as SKYBORG, Resolute Sentry, Autonomous Air Combat Operations (AACO), and Autonomous Attributable Aircraft Experiment.