Anomaly Detection Using Adaptive Suppression
Artur Grigorev, Oksana Severiukhina, Ivan Derevitskii · Procedia Computer Science · 2019
The paper presents an unsupervised method for anomaly detection named Adaptive Suppression (AS). The method relies on distance metrics between data points and receives hypothesis about percent of anomalies in dataset. Data points having initial accumulator value, “suppress” each other based on distance. Data points with high retained accumulator values considered as anomalous. We tested the method using anomaly video datasets (QMUL and UCSD). We extract scene dynamics using grid of short-time trackers (TrackGrid) and apply Adaptive Suppression to every TrackLoop (element of TrackGrid) separately, and then aggregate values using maximum. Using proposed method, we obtain competitive results: 0.746 AUC on QMUL, 0.7 on UCSD Ped1, 0.82 AUC on UCSD Ped2. Also, we apply Adaptive Suppression to CIFAR-10 dataset over space of reduced dimensionality (using UMAP) to provide an example of applicability of AS to image data. Possibly, AS has simple implementation and can be applied to any kind of data for which distance (or similarity metrics) exists, allowing to estimate anomaly/uniqueness of texts, images and other types of datasets using Deep Learning methods.