[3Xin2] Poster session 1

[3Xin2-52] Vulnerability detection in network environments by multi-process pseudo attacks with reinforcement learning

〇Koki Takeshita¹, Takuma Shibahara¹ (1.Hitachi, Ltd.)

Keywords:Reinforcement learning, Cybersecurity

In recent years, the number of cyber-attacks has greatly increased with the spread of new working styles due to the COVID-19 pandemic, an increase in remote work, and the digital transformation of organizations including corporations and hospitals, the number of cyber-attacks has greatly increased. Accordingly, the demand for vulnerability detection technology has increasing. Fuzzing tests and DDoS drills with input generated based on rules, have been widely used for vulnerability detection. Alternatively, in recent years, there have been researches on vulnerability detection by pseudo attack with reinforcement learning. With the advantage of reinforcement learning that is able to efficiently explore effective actions from a vast action space, it is able to discover pseudo attacks that exploit system vulnerabilities. In contrast, pseudo attacks with reinforcement learning often results in longer generation time compared to rule-based methods, thereby preventing the replication of multi-vector DDoS attacks that necessitate the creation of multiple pseudo attacks within a limited timeframe. Therefore, in this method, multi-agent we apply reinforcement learning to train multi-process pseudo attacks, achieving pseudo attacks by multiple agents.