This project applies artificial-intelligence (AI) techniques to proactive vulnerability assessment (VA) in computer networks. Current library-based approach to VA does not allow the exploitation of vulnerabilities outside the library of the known attack characteristics. This proposal takes the first step toward the next generation of proactive VA software by developing advanced AI techniques that learn to attack a computer network, and hence discover its vulnerabilities and weaknesses before these weaknesses are exploited. The initial work casts VA within the framework of reinforcement learning (RL) as this approach has demonstrated previous successes for other networking problems. RL researchers study algorithms for learning high reward strategies for one or more agents based on reward signals received while interacting with an environment. For VA, the environment corresponds to a specific computer network; the reward signal provides positive reward for activity that is detrimental to a network and negative reward for activity that is detected as malicious (hence cannot harm the network). The strategy discovered by RL gives a method for one or more agents to attack the network without being detected. In this proposal, the focus is on using RL techniques to discover VA in Peer-to-Peer networks. The broader impact of this project will include bridging the gap between AI and network research communities for the purpose of providing a strong defense against network attacks. Research results will be disseminated through a website at www.eecs.orst.edu/~thinhq/research/AI_Security/index.html. The research project will provide a hands-on research and learning environment for students and contribute to the development of Cyber Trust workforce.
