OneFlow: One-Class Flow for Anomaly Detection Based on a Minimal Volume Region.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Anomaly detection is crucial for identifying unusual patterns in data.
• Existing density-based methods can be sensitive to the structure of outliers.
• One-class classification aims to model normal data to detect deviations.

Purpose of the Study:

• To introduce OneFlow, a flow-based one-class classifier for anomaly detection.
• To develop a method that is robust to the structure of outliers.
• To create a classifier that defines a minimal volume bounding region.

Main Methods:

• Utilizes flow models combined with a Bernstein quantile estimator.
• Employs a training approach where gradients propagate near the decision boundary, similar to Support Vector Machines (SVM).
• Focuses on learning a parametric form for the bounding region.

Main Results:

• OneFlow demonstrates independence from the structure of outliers.
• The classifier effectively identifies a minimal volume bounding region.
• Achieves superior performance compared to existing methods on real-world anomaly detection tasks.

Conclusions:

• OneFlow offers a robust and effective approach to anomaly detection.
• The parametric bounding region is suitable for diverse applications, including 3D point cloud analysis.
• The method provides a significant advancement over traditional density-based techniques.