SADDE: Semi-supervised Anomaly Detection with Dependable Explanations

원문 초록

Semi-supervised learning holds a pivotal position in anomaly detection applications, yet identifying anomaly patterns with a limited number of labeled samples poses a significant challenge. Furthermore, the absence of interpretability poses major obstacles to the practical adoption of semi-supervised frameworks. The majority of existing interpretation techniques are tailored for supervised/unsupervised frameworks or non-security domains, falling short in providing dependable interpretations. In this research paper, we introduce SADDE, a general framework designed to accomplish two primary objectives: (1) to render the anomaly detection process interpretable and enhance the credibility of interpretation outcomes, and (2) to assign high-confidence pseudo labels to unlabeled samples, thereby boosting the performance of anomaly detection systems when supervised data is scarce. To achieve the first objective, we devise a cutting-edge interpretation method that utilizes both global and local interpreters to furnish trustworthy explanations. For the second objective, we conceptualize a novel two-stage semi-supervised learning framework tailored for network anomaly detection, ensuring that the model predictions of both stages align with specific constraints. We apply SADDE to two illustrative network anomaly detection tasks and conduct extensive evaluations in comparison with notable prior works. The experimental findings underscore that SADDE is capable of delivering precise detection results alongside dependable interpretations for semi-supervised network anomaly detection systems. The source code for SADDE is accessible at: https://github.com/M-Code-Space/SADDE.