Adversarial Machine Learning in Cybersecurity: Defending Against Evasive Threats in Distributed Systems

This paper investigates the security issues raised by adversarial evasion attacks against malware detection systems that use machine learning. We evaluate methods used by malware authors to create evasive variations and countermeasures used by defenders to keep model robustness. Our study demonstrates a continuous arms race: although well designed adversarial malware can greatly lower detection rates, adaptive defences such as adversarial training and feature-space regularization can greatly recover performance. We provide empirical results from recent studies to show the practical influence of both attack and defense techniques as well as mathematical definitions of each. For security experts trying to create strong AI-based malware detection systems in the face of ever complex threats, this work offers valuable insights.