A Study on Distance Measure for Effective Anomaly Detection using AutoEncoder

Abstract

Anomaly detection is a popular application in various areas. One challenging issue is to build an anomaly detection model using normal data because collecting potential abnormal data is quite difficult. In this paper, we build an anomaly detection model using just normal data based on adversarial autoencoder for acoustic data. After extracting features using the trained model, we apply a distance-based method for calculating a threshold to be used for anomaly detection. In particular, we propose a method for reflecting differences in dimensions in calculating distance. Through experiments, we show that the proposed dimension-aware distance measure improves anomaly detection accuracy by up to 7% compared to existing distance measure methods.