Defending network traffic attack with distributed multi-agent reinforcement learning

Abstract

The DDoS attack is a serious security problem in today’s Internet. To defend DDoS attacks, distributed routers need to cooperate to guarantee servers’ safety. Existing rule-based methods are not ideal as these methods adjust a threshold, which lacks flexibility. The distributed DDoS defense problem can also be viewed as a multi-agent Markov decision process (MAMDP), where multiple agents intelligently coordinate to achieve a team goal. However, agents can hardly defend DDoS attacks without communication. In this paper, we propose a multi-agent reinforcement learning method for distributed routers to defend DDoS attacks via compressed communication. Our method named ComDDPG outperforms existing rule-based methods and independent reinforcement learning methods under diverse attack scenarios even with communication delay.