Fuzzy Clustering based Anomaly Detection for Distributed Multi-view Data

Anomaly detection aims to identify the abnormal instances, whose behavior deviates significantly from the others. Nowadays owing to the existence of diverse data generation sources, different attributes of the same instances may be located on distributed parties forming a multi-view dataset. Thus multi-view anomaly detection has become a key task to discover outliers across various views. Traditionally, to perform multi-view anomaly detection, one needs to centralize data instances from all views into a single machine. However, in many real-world scenarios, it is impractical to send data from diverse views to a master machine due to the privacy issues. Inspired by this, we propose a fuzzy clustering based distributed approach for multi-view anomaly detection that simultaneously learns a membership degree matrix for each view and then detects anomalies for all parties. Specifically, we first introduce a combined fuzzy c-means clustering method for multi-view data and then design an anomaly measurement criterion to quantify the abnormal score from membership degree matrix. To solve the proposed model, a protocol is provided to unify all parties performing a well-designed optimization in an iterative way. Experiments on three datasets with different anomaly settings demonstrate the effectiveness of our approach.