Adversarial Data Augmentation for HMM-Based Anomaly Detection.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Anomaly detection in physical systems is challenging due to the unpredictability of all possible anomalies.
• Existing methods often struggle with multivariate time series data and one-class classification scenarios.
• Adversarial data augmentation techniques are predominantly explored in image recognition, not time series analysis.

Purpose of the Study:

• To develop and evaluate a novel data augmentation and retraining approach for anomaly detection using adversarial learning.
• To improve the performance and robustness of anomaly detection systems, particularly for multivariate time series.
• To address the limitations of current methods in handling unknown anomalies in physical systems.

Main Methods:

• A method for generating adversarial examples tailored for anomaly detectors based on Hidden Markov Models (HMMs) was defined.
• A data augmentation and retraining technique was developed, utilizing these generated adversarial examples.
• The approach was evaluated on four distinct datasets, focusing on multivariate time series and one-class classification.

Main Results:

• The proposed adversarial data augmentation and retraining approach demonstrated statistically significant improvements in anomaly detection performance.
• The method significantly enhanced the robustness of anomaly detection systems against adversarial attacks.
• Performance gains were observed across multiple datasets, validating the effectiveness of the technique.

Conclusions:

• Adversarial learning, combined with HMMs, offers a powerful strategy for enhancing anomaly detection in physical systems.
• The developed technique effectively improves detection accuracy and resilience to novel, adversarial threats.
• This work advances the state-of-the-art in anomaly detection for multivariate time series within a one-class classification framework.