PRAAD: Pseudo representation adversarial learning for unsupervised anomaly detection

Abstract

As one of the typical means of anomaly detection, unsupervised reconstruction-based anomaly detection methods usually extract the normal representations and utilize the reconstruction error to detect the anomalies. The main framework is autoencoder. If the autoencoder has strong generalization ability, the anomalies could also be well reconstructed, resulting in model misjudgment. Therefore, we propose a Pseudo Representation Adversarial learning model for unsupervised Anomaly Detection (PRAAD). Specifically, we design a pseudo-representation-based data augmentation strategy to enrich latent distribution for capturing additional normal patterns through an adversarial learning strategy. Based on this, PRAAD could improve the confidence of normal reconstruction rather than abnormal reconstruction. Finally, we additionally consider the distance of the sample to the latent distribution to synthesize the anomaly score. Experimental results on real image and cybersecurity datasets show that PRAAD outperforms the state-of-the-art baselines.