Unsupervised Anomaly Detection by Robust Density Estimation.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Density estimation is crucial for unsupervised anomaly detection.
• Anomalies in training data can compromise sophisticated density estimation methods, including deep neural networks.
• Existing flow-based models are susceptible to data corruption by anomalies.

Purpose of the Study:

• To develop a robust deep density estimation framework for unsupervised anomaly detection.
• To enhance the accuracy and reliability of anomaly detection in the presence of noisy training data.
• To overcome the limitations of current methods when faced with anomalous training samples.

Main Methods:

• Proposing RobustRealNVP, a novel deep density estimation framework.
• Implementing a strategy to discard low-density data points during optimization to prevent corruption.
• Incorporating Lipschitz regularization to ensure a smooth estimated density function.

Main Results:

• Demonstrated theoretical and empirical robustness of the algorithm against anomalies in training data.
• RobustRealNVP significantly outperforms existing state-of-the-art unsupervised anomaly detection methods.
• The proposed method effectively prevents anomalies from negatively impacting the density estimation process.

Conclusions:

• RobustRealNVP offers a robust solution for unsupervised anomaly detection, even with contaminated training data.
• The framework enhances the performance of deep density estimation by mitigating the impact of anomalies.
• This work advances the field of anomaly detection by providing a more reliable and accurate approach.