Adversarially Reweighted Sequence Anomaly Detection With Limited Log Data

Abstract

In the realm of safeguarding digital systems, the ability to detect anomalies in log sequences is paramount, with applications spanning cybersecurity, network surveillance, and financial transaction monitoring. This thesis presents AdvSVDD, a sophisticated deep learning model designed for sequence anomaly detection. Built upon the foundation of Deep Support Vector Data Description (Deep SVDD), AdvSVDD stands out by incorporating Adversarial Reweighted Learning (ARL) to enhance its performance, particularly when confronted with limited training data. By leveraging the Deep SVDD technique to map normal log sequences into a hypersphere and harnessing the amplification effects of Adversarial Reweighted Learning, AdvSVDD demonstrates remarkable efficacy in anomaly detection. Empirical evaluations on the BlueGene/L (BG/L) and Thunderbird supercomputer datasets showcase AdvSVDD’s superiority over conventional machine learning and deep learning approaches, including the foundational Deep SVDD framework. Performance metrics such as Precision, Recall, F1-Score, ROC AUC, and PR AUC attest to its proficiency. Furthermore, the study emphasizes AdvSVDD’s effectiveness under constrained training data and offers valuable insights into the role of adversarial component has in the enhancement of anomaly detection.