Supervised Anomaly Detection via Conditional Generative Adversarial Network and Ensemble Active Learning.

Area of Science:

• Machine Intelligence
• Data Science

Background:

• Anomaly detection is crucial but challenging due to rare, imbalanced anomaly classes.
• Traditional methods like unsupervised detectors are suboptimal, and supervised models risk bias towards normal data.

Purpose of the Study:

• To develop a novel supervised anomaly detection method overcoming limitations of existing approaches.
• To introduce the Ensemble Active Learning Generative Adversarial Network (EAL-GAN) for enhanced anomaly detection.

Main Methods:

• Utilized a conditional Generative Adversarial Network (GAN) with a unique one-generator-multiple-discriminator architecture.
• Implemented anomaly detection via an auxiliary classifier within the discriminator.
• Incorporated an ensemble learning loss function and an active learning algorithm to address class imbalance and reduce labeling costs.

Main Results:

• The proposed EAL-GAN consistently outperformed state-of-the-art (SOTA) anomaly detection methods.
• Demonstrated significant improvements in detecting rare and imbalanced anomaly classes.

Conclusions:

• EAL-GAN offers a robust and efficient solution for supervised anomaly detection.
• The method effectively handles class imbalance and reduces the need for extensive data labeling.