Hiding clusters in adversarial settings

摘要: In adversarial settings, records associated with those who want to conceal their existence or activities tend be unusual because of illicit status; but not too efforts make them as normal possible. Clusters such will single outliers even outlying clusters, rather small clusters on the fringes clusters. Such structures are undetectable by many mainstream clustering algorithms, for example based distance and convexity. We show that more sophisticated algorithms easily subverted addition only a few carefully chosen records. Robust in settings require development tailored this domain.