Adversarial Reinforcement Learning in a Cyber Security Simulation

Elderman, R. and Pater, L.J.J. and Thie, A.S. (2016) Adversarial Reinforcement Learning in a Cyber Security Simulation. Bachelor's Thesis, Artificial Intelligence.

Preview	Text AI_BA_2016_RICHARDELDERMAN.pdf - Published Version Download (528kB) | Preview
	Text Toestemming.pdf - Other Restricted to Backend only Download (569kB)

Abstract

In this thesis we focus on the problem of securing assets and data in a computer network. After building our own security game, we train attacking and defending agents against each other. The security game is modelled as a sequential decision making problem for the attacker and defender. The game is simulated as an extensive form game with incomplete information and stochastic elements. By using neural networks, Monte Carlo learning and Q-learning we pit various reinforcement learning techniques against each other to examine their effectiveness against learning opponents. We found Q-learning and Monte Carlo learning with epsilon-greedy exploration to be most effective in performing the defender role and different attacker algorithms based on Monte Carlo learning to be most effective in attacking.

Item Type:	Thesis (Bachelor's Thesis)
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	15 Feb 2018 08:24
Last Modified:	15 Feb 2018 08:25
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/14523

Actions (login required)

View Item