Center-Aware Adversarial Autoencoder for Anomaly Detection.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Anomaly detection commonly relies on subspace learning, focusing on subspace compactness.
• Existing methods often fail to adapt normal areas by not considering individual sample anomaly levels.
• This limitation hinders effective isolation of normal from anomalous data.

Purpose of the Study:

• To propose a Center-Aware Adversarial Autoencoder (CA-AAE) for improved anomaly detection.
• To enhance subspace representations for better compactness and discriminability.
• To adaptively adjust normal areas by considering the anomaly level of normal samples.

Main Methods:

• Developed a CA-AAE integrating anomaly-level description and feature learning via presubspace and postsubspace.
• Implemented adversarial learning in presubspace to impose a toward-center prior distribution.
• Introduced a center-aware strategy in postsubspace with a weighted center for adaptive normal area adjustment.

Main Results:

• The CA-AAE method achieves more compact and discriminative subspace representations.
• Anomaly levels of normal samples are described probabilistically.
• Numerical experiments demonstrate the effectiveness and advantages of the proposed CA-AAE.

Conclusions:

• The proposed CA-AAE effectively enhances anomaly detection through adaptive subspace learning.
• Integrating anomaly level description and a center-aware strategy improves the isolation of normal and anomalous data.
• CA-AAE offers a promising approach for anomaly detection tasks.