Can a good offense be a good defense? Vulnerability testing of anomaly detectors through an artificial arms race

Intrusion detection systems, which aim to protect our IT infrastructure are not infallible. Attackers take advantage of detector vulnerabilities and weaknesses to evade detection, hence hindering the effectiveness of the detectors. To do so, attackers generate evasion attacks which can eliminate or minimize the detection while successfully achieving the attacker’s goals. This work proposes an artificial arms race between an automated ‘white-hat’ attacker and various anomaly detectors for the purpose of identifying detector weaknesses. The proposed arms rac aims to automate the vulnerability testing of the anomaly detectors so that the security experts can be more proactive in eliminating detector vulnerabilities.