Abstract

Outlier detection methods typically assume clean and legitimate data streams. However, adversaries may attempt to influence data which in turn may impact outlier designations. This paper presents a decision theoretic approach for outlier detection in adversarial environments. Proposed adversarial risk analysis based framework allows incomplete information and adversarial perturbations on the data inputs. We solve the adversary’s poisoning decision problem where he manipulates batch data inputted into outlier detection methods. We discuss potential defender strategies to improve the security of existing frameworks.