Mathematical Modelling of Cyber-Attackers vs Defenders using AI/ML (available)

Starting Date: Summer 2019
Duration: 10-12 weeks
Time commitment: 20 hours a week
Prerequisites: Strong math background, knowledge of Python or similar language, some machine Learning notions might be useful

Machine learning is a popular approach to signature-less malware detection because it can generalize to new (unseen) malware families. Some recent works have proposed the use of AI/ML-powered malware to bypass machine learning anti-malware systems (for instance, adversarial machine learning).

The goal of the project is to model the system of malware vs anti-malware systems as two opponents using various AI/ML strategies to bypass the other side, such as adversarial machine learning. We would like to model this system using a mathematical (probabilistic) model to understand how the real system evolves, e.g. whether it converges to a stable state (and who benefits from this state: attackers or defenders?) or to a never-ending game, or whether the convergence (if any) depends on some initial assumptions.

Some references:

https://arxiv.org/abs/1801.08917
https://github.com/endgameinc/gym-malware
https://i.blackhat.com/us-18/Thu-August-9/us-18-Kirat-DeepLocker-Concealing-Targeted-Attacks-with-AI-Locksmithing.pdf
https://dl.acm.org/citation.cfm?id=3150378
https://arxiv.org/pdf/1811.01190.pdf
https://github.com/a0rtega/pafish
https://arxiv.org/abs/1712.03141