ASSURED INFORMATION SECURITY, INC. — Department of Defense SBIR Phase I: N193-A01

Description

AIS proposes studying adversarial attacks on RL-agents within the StarCraft II AI-learning environment. StarCraft II provides a rich environment to test RL agents and acts as a foundational surrogate for more general problems such as resource management and military strategy. The objective of this effort is to develop and evaluate methods that disrupt an RL-agents ability to learn and perform effectively through manipulation of the environment (triggers), rewards (false signals), and enemy actions (mis-leading intelligence) that an adversary can feasibly change. These attack vectors are unique to RL, distinct from the data manipulation attacks prevalent in supervised learning research, and more likely to be exploited by adversaries. AIS will develop and optimize a series of adversarial attacks based upon these attacks on existing RL-agents for StarCraft II and evaluate their ability to disrupt or influence the agents’ performance and actions.