Distributed reinforcement learning for network intrusion response

• Main Author: Malialis, Kleanthis
• Other Authors: Kudenko, Daniel
• Published: University of York 2014
• Subjects: 004
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.638996

The increasing adoption of technologies and the exponential growth of networks has made the area of information technology an integral part of our lives, where network security plays a vital role. One of the most serious threats in the current Internet is posed by distributed denial of service (DDoS) attacks, which target the availability of the victim system. Such an attack is designed to exhaust a server's resources or congest a network's infrastructure, and therefore renders the victim incapable of providing services to its legitimate users or customers. To tackle the distributed nature of these attacks, a distributed and coordinated defence mechanism is necessary, where many defensive nodes, across different locations cooperate in order to stop or reduce the flood. This thesis investigates the applicability of distributed reinforcement learning to intrusion response, specifically, DDoS response. We propose a novel approach to respond to DDoS attacks called Multiagent Router Throttling. Multiagent Router Throttling provides an agent-based distributed response to the DDoS problem, where multiple reinforcement learning agents are installed on a set of routers and learn to rate-limit or throttle traffic towards a victim server. One of the novel characteristics of the proposed approach is that it has a decentralised architecture and provides a decentralised coordinated response to the DDoS problem, thus being resilient to the attacks themselves. Scalability constitutes a critical aspect of a defence system since a non-scalable mechanism will never be considered, let alone adopted, for wide deployment by a company or organisation. We propose Coordinated Team Learning (CTL) which is a novel design to the original Multiagent Router Throttling approach based on the divide-and-conquer paradigm, that uses task decomposition and coordinated team rewards. To better scale-up CTL is combined with a form of reward shaping. The scalability of the proposed system is successfully demonstrated in experiments involving up to 1000 reinforcement learning agents. The significant improvements on scalability and learning speed lay the foundations for a potential real-world deployment.