Multiple Level Action Embedding for Penetration Testing

Abstract

Penetration Testing (PT) is one of the most effective and widely used methods to increase the defence of a system by looking for potential vulnerabilities. Reinforcement learning (RL), a powerful type of machine learning in self-decision making, is demonstrated to be applicable in PT to increase automation as well as reduce implementation costs. However, RL algorithms are still having difficulty on PT problems which have large network size and high complexity. This paper proposes a multiple level action embedding applied with Wolpertinger architect (WA) to enhance the accuracy and performance of the RL, especially in large and complicated environments. The main purpose of the action embedding is to be able to represent the elements in the RL action space as an n-dimensional vector while preserving their properties and accurately representing the relationship between them. Experiments are conducted to evaluate the logical accuracy of the action embedding. The deep Q-network algorithm is also used as a baseline for comparing with WA using the multiple level action embedding.