Anomaly Detection Using Complete Cycle Consistent Generative Adversarial Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Traditional anomaly detection methods struggle with high variance in accuracy across different anomaly types.
• Existing techniques are often ineffective in real-world scenarios with diverse data distributions.

Purpose of the Study:

• To develop a robust adversarial method for anomaly detection using generative adversarial neural networks (GANs).
• To improve detection precision and overcome limitations of traditional anomaly detection approaches.

Main Methods:

• Leveraging generative adversarial neural networks (GANs) with cycle consistency in reconstruction error.
• Introducing an innovative information flow and a new discriminator to optimize training dynamics.
• Employing a supplementary distribution in the input space to guide reconstructions towards normal data distribution.
• Developing two unique anomaly scoring mechanisms for augmented detection capabilities.

Main Results:

• The proposed model demonstrates superior performance compared to one-class anomaly detection benchmarks.
• Comprehensive evaluations on six diverse datasets confirm the model's effectiveness.
• The method successfully isolates anomalous instances and enhances detection precision.

Conclusions:

• The developed adversarial method offers a robust solution for real-world anomaly detection.
• The innovative training procedure and scoring mechanisms significantly improve detection capabilities.
• The open-source implementation facilitates further research and application in the academic community.