AGAD: Adversarial Generative Anomaly Detection
This addresses the challenge of robust anomaly detection for applications where anomaly data is scarce, representing an incremental advancement by generating pseudo-anomalies to enhance existing methods.
The paper tackles the problem of anomaly detection with limited anomaly data by proposing AGAD, a self-contrast-based method that generates pseudo-anomaly data from normal examples, achieving significant improvements in both supervised and semi-supervised scenarios and boosting detection accuracy with no more than 5% anomalous training data.
Anomaly detection suffered from the lack of anomalies due to the diversity of abnormalities and the difficulties of obtaining large-scale anomaly data. Semi-supervised anomaly detection methods are often used to solely leverage normal data to detect abnormalities that deviated from the learnt normality distributions. Meanwhile, given the fact that limited anomaly data can be obtained with a minor cost in practice, some researches also investigated anomaly detection methods under supervised scenarios with limited anomaly data. In order to address the lack of abnormal data for robust anomaly detection, we propose Adversarial Generative Anomaly Detection (AGAD), a self-contrast-based anomaly detection paradigm that learns to detect anomalies by generating \textit{contextual adversarial information} from the massive normal examples. Essentially, our method generates pseudo-anomaly data for both supervised and semi-supervised anomaly detection scenarios. Extensive experiments are carried out on multiple benchmark datasets and real-world datasets, the results show significant improvement in both supervised and semi-supervised scenarios. Importantly, our approach is data-efficient that can boost up the detection accuracy with no more than 5% anomalous training data.