Unsupervised Adversarial Anomaly Detection using One-Class Support Vector Machines

Anomaly detection discovers regular patterns in unlabeled data and identifies the non-conforming data points, which in some cases are the result of malicious attacks by adversaries. Learners such as One-Class Support Vector Machines (OCSVMs) have been successfully used in anomaly detection, yet their performance may degrade significantly in adversarial conditions such as integrity attacks. This work focuses on integrity attacks, where the adversary distorts the training data in order to successfully avoid detection during evaluation. This paper presents a unique combination of anomaly detection using (1) OCSVMs in the presence of adversaries who distort training data in a targeted manner and (2) nonlinear randomized kernel methods, which facilitate computational and conceptual simplification through dimension reduction. We theoretically analyze the effects of adversarial distortions on the separating margin of OCSVMs and provide supporting empirical evidence. The proposed approach introduces a layer of uncertainty on top of the OCSVM learner, making it challenging for the adversary to guess the specific configuration of the learner.