Generative adversarial synthetic neighbors-based unsupervised anomaly detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Anomaly detection is vital for system stability, financial security, and network integrity.
• Existing Generative Adversarial Networks (GANs) methods for anomaly detection often require labeled data or struggle with efficiency and generalization on complex data distributions.

Purpose of the Study:

• To address limitations in current unsupervised anomaly detection methods.
• To introduce a novel Generative Adversarial Synthetic Neighbors (GASN) based unsupervised anomaly detection method.

Main Methods:

• GASN integrates Generative Adversarial Networks (GANs) with neighborhood analysis for a two-stage detection process.
• The first stage involves training GANs to model normal data distributions and generate synthetic normal samples.
• The second stage uses neighborhood analysis to compare original and synthetic data, calculating an anomaly factor for each object.

Main Results:

• The proposed GASN method demonstrated superior performance in anomaly detection.
• Experiments on twelve public datasets showed GASN improved the Area Under the Curve (AUC) by 9.93% compared to the second-best method.
• GASN effectively detects subtle anomalies by leveraging synthetic data and neighborhood comparisons.

Conclusions:

• GASN offers an effective unsupervised approach to anomaly detection, outperforming existing state-of-the-art methods.
• The method shows promise for applications requiring robust anomaly detection in complex datasets.
• GASN enhances anomaly detection by improving computational efficiency and generalization capabilities.