Dungeons, Dragons, and Data Breaches: Analyzing AI Attacks on Various Network Configurations

This chapter studies the usage of artificial intelligence in network analysis when applied to the problem of attempting to identify security vulnerabilities. This chapter demonstrates how AI is capable of more efficient and detailed analysis than a human cyber-security individual by using the CyberBattleSim simulation program developed by Microsoft, which utilizes Deep-Q-learning algorithms to train the AI. During the investigation conducted in this report, several environments are studied: three networks of varying size based on the chain network pattern provided by the initial program, three networks based on the toy network pattern, and a custom network designed to examine the AI’s capabilities in handling a unique environment involving honeypots and two-factor authentication. The AI was determined to be adept at learning from one episode to the next, accumulating rewards more quickly, and using less time to take control of the network. Based on the research conducted, multiple modifications to improve the system are suggested.