Anomaly Detection Based on Zero-Shot Outlier Synthesis and Hierarchical Feature Distillation

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Anomaly detection is challenged by highly imbalanced datasets where anomalies are rare.
• Existing synthetic data generation methods require expressive representations for high-quality anomaly synthesis.
• Zero-shot anomaly generation necessitates robust feature descriptors and generative models.

Purpose of the Study:

• To propose a two-level hierarchical latent space representation for robust feature distillation.
• To enable zero-shot anomaly generation and improve outlier detection performance.
• To train effective binary classifiers using synthetically generated negative samples.

Main Methods:

• Utilized autoencoders (AEs) for initial feature descriptor distillation.
• Employed variational autoencoders (VAEs) to create robust representations from a variational family of distributions.
• Developed a hierarchical structure for feature distillation and fusion.
• Selected latent distributions on the outskirts of training data for synthetic-outlier generation.

Main Results:

• The hierarchical structure yields robust and generalizable representations.
• Successfully synthesized pseudo outlier samples without prior exposure to real outliers.
• Trained effective binary classifiers for true outlier detection using synthetic data.
• Demonstrated strong performance on multiple anomaly detection benchmarks.

Conclusions:

• The proposed hierarchical latent space effectively addresses the challenge of imbalanced data in anomaly detection.
• Zero-shot anomaly generation is feasible and beneficial for training robust classifiers.
• This approach offers a viable solution for scenarios lacking sufficient or defined outlier data.