Autoencoders with exponential deviation loss for weakly supervised anomaly detection

Abstract

Listen

Translate article

Weakly supervised anomaly detection aims to detect anomalies using a small number of labeled anomalies and a large amount of unlabeled data. However, existing methods have limitations: unsuitable setting of the anomaly threshold, using an inefficient loss function, and being vulnerable to contaminated data. To address these limitations, this paper proposes a novel framework called the Exponential Deviation Autoencoder (EDAE), which consists of two stages. In the first stage, EDAE pre-trains an autoencoder (AE) to learn a compressed representation of the input data and estimates the anomaly score distribution of the training data to determine an appropriate anomaly threshold. In the second stage, EDAE fine-tunes the AE with a novel Exponential Deviation Loss (EDL) function that provides continuous and nonlinear penalties according to anomaly scores and enables more effective training using labeled anomalies. EDAE also uses batch sampling based on empirical distribution to create batches of data that are more robust to contaminated data. We conduct extensive experiments on various datasets and show that EDAE outperforms state-of-the-art weakly supervised methods with up to a 26% improvement in accuracy.