Machine learning for offensive cyber operations

Abstract

Cybersecurity with machine learning has received widespread attention in education, research, and innovation in both the private and public sectors. Unfortunately, while essential for strong cyber security, offensive cyber operations with machine learning have seen significantly less innovation, at least in open academic literature. This thesis's contribution to the field of offensive cyber operations with machine learning can naturally be divided into the following: 1. Algorithmic cryptanalysis with machine learning 2. SQL injection with machine learning The historical cipher Enigma's plugboard is shown to be susceptible to an attack powered by the machine learning technique Genetic Algorithms, being broken far faster than any earlier attack. Modern ciphers are naturally more robust than historical ciphers. The cryptographic algorithm ASCON is still secure, but the novel machine learning technique, The Phantom Gradient Attack, is shown to attack many of its subfunctions successfully. OWASP's top 10 had SQL injections as the number one web vulnerability in 2017; in 2021, it was number three. This thesis highlights the possibility of automated SQL injection exploitation and identification with reinforcement learning for accelerated penetration testing. The reinforcement learning agent can exploit all 5 SQL injection archetypes, distinguish between them, and determine whether or not the website is vulnerable.