NFAD: fixing anomaly detection using normalizing flows.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Anomaly detection is crucial across various fields, including fraud detection and medical diagnosis.
• Conventional one-class classification methods often assume separability and ignore available anomaly samples.
• Real-world anomaly detection often involves limited, imbalanced anomaly data where separability may not hold.

Purpose of the Study:

• To develop a novel, model-agnostic training procedure for anomaly detection that incorporates known anomaly samples.
• To address the limitations of existing methods that do not utilize available, albeit scarce, anomalous data.
• To improve the performance of anomaly detection models in practical settings with imbalanced datasets.

Main Methods:

• Reformulated one-class classification as a binary classification problem.
• Generated pseudo-anomalous samples from low-density regions of a normalizing flow model.
• Developed a model-agnostic training procedure to integrate known anomalies into arbitrary classifiers.

Main Results:

• The proposed approach demonstrated comparable performance on traditional one-class anomaly detection tasks.
• Achieved comparable or superior results on tasks with varying amounts of known anomalies.
• Successfully incorporated known anomalies into the training of various anomaly detection models.

Conclusions:

• The novel training procedure effectively utilizes known anomaly samples, even when scarce.
• The method offers a flexible and powerful way to enhance anomaly detection models.
• This approach advances anomaly detection by better handling real-world data characteristics.