Stabilizing Adversarially Learned One-Class Novelty Detection Using Pseudo Anomalies.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Anomaly detection models often use reconstruction or classification loss.
• Training these models is challenging due to the lack of anomaly examples and training instability.
• Adversarial training leads to performance fluctuations, complicating optimal training termination.

Purpose of the Study:

• To propose a robust anomaly detection framework that addresses training instability.
• To redefine the discriminator's role from real/fake classification to distinguishing reconstruction quality.
• To develop an efficient criterion for optimal model training termination.

Main Methods:

• A novel discriminator role is proposed: distinguishing good vs. bad quality reconstructions.
• Good and bad reconstructions are generated using current and old states of the same generator.
• The discriminator is trained to detect subtle distortions in reconstructions of anomalous data.
• An efficient, generic criterion is introduced for stopping model training.

Main Results:

• The proposed framework demonstrates robust anomaly detection across diverse datasets and domains.
• Experiments cover image, video, medical diagnosis, and network security anomaly detection.
• The approach achieves excellent performance, overcoming limitations of previous methods.

Conclusions:

• The developed framework offers a stable and effective solution for anomaly detection.
• Transforming the discriminator's role and using quality-based examples enhances model performance.
• The efficient training termination criterion ensures elevated and consistent results.