AGAD: Adversarial Generative Anomaly Detection

Jian Shi,Ni Zhang

2023-04-09

Abstract: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.

Computer Vision and Pattern Recognition,Machine Learning

What problem does this paper attempt to address?

The paper aims to address the issue of data scarcity in anomaly detection, particularly the challenge of the rarity and diversity of anomalous samples. Specifically, the study proposes a new method called Adversarial Generative Anomaly Detection (AGAD), which is a self-contrastive anomaly detection paradigm. The core idea is to generate contextual adversarial information from a large number of normal examples to learn to detect anomalies.

The main contributions of AGAD include:
1. **Introducing contextual adversarial information**: By generating pseudo-anomalous data to better learn the discriminative features between normal and anomalous data.
2. **Proposing a simple yet effective anomaly detection paradigm**: This method can unify supervised and semi-supervised anomaly detection schemes and can significantly improve performance with a small amount of anomalous data.
3. **Data efficiency**: The proposed method can effectively enhance model performance even with less than 5% of anomalous images.

This method is applicable to both supervised and semi-supervised anomaly detection scenarios. Experimental results show that AGAD significantly outperforms existing techniques on various benchmark datasets, whether in semi-supervised or limited-supervised anomaly detection tasks.

In short, AGAD aims to solve the problem of data insufficiency in anomaly detection through adversarial generation, especially in practical applications where anomalous data is scarce, and it can achieve efficient anomaly detection with limited data.