Sandia National Laboratories Richmond, Virginia, United States
Physical security of nuclear facilities is an important regulatory goal, but can be expensive. The current state of practice for designing these systems involves manual design by human experts. This approach is effective, but expensive, time consuming, biased towards historical examples, and lacks any effectiveness guarantees for threats outside regulatory guidelines. This work is aimed at developing a substantial leap forward in physical security design through first designing an adversarial machine learning agent learns to exploit existing approaches. Next, a designer agent will be trained to create candidate designs, based on user criteria such as price or footprint, that should be subsequently refined by human experts. These agents could ultimately be trained together through self-play, but such an approach has several theoretical challenges.
This talk and poster will focus on our efforts in developing the adversarial agent. We will show how we trained an agent from scratch that learned the optimal exploitation behavior for a physical protection system around a notional nuclear facility. The agent we developed was not provided any explicit instructions and had to learn the rules and goals of the ``game'' entirely through self-play. Later, we will show that the agent struggles to generalize to new facilities without additional training and we will discuss our ongoing efforts to combat this limitation.
