Classification of Malicious Network Activity

As more and more vital services today (e.g. email, Facebook, quantitative trading) depend on machine learning algorithms, there is a greater incentive than ever for adversaries to manipulate these algorithms for malicious ends (e.g. spam, identity theft, cyberattacks). The field of adversarial learning has arisen out of a need to design learning algorithms that are robust to these sort of disruptive manipulations. Spam filtering has often been cited as the classic adversarial machine learning problem [5]. As filtering methods grow more accurate, spammers employ more and more sophisticated methods to work around them (e.g. deliberate misspellings of keywords, etc.), resulting in an ongoing “arms race”. A more grave example, however, that illustrates the need for robust learning methods is the network intrusion problem, where potentially sensitive, confidential, or damaging information is compromised through sophisticated cyberattacks on computer networks. Cybercrime and cyberwarfare are real problems [2, 6] that advances in adversarial machine learning may mitigate.