Deep Learning for Collective Anomaly Detection

Anomaly detection has been a cornerstone of research across diverse disciplines, remaining a critical and evolving field of study for several decades. While several approaches, including deep learning (DL), have been explored to design general solutions for anomaly detection, there is a lack of structured survey articles on the collective (or group) anomaly detection problem (CAD). This article fills this gap and presents the first comprehensive review dedicated exclusively to DL-based methods for CAD. A two-level taxonomy is proposed to categorize CAD methods based on their underlying algorithms into generative, discriminative, or hybrid, and the methods are further classified according to their DL architecture. The literature review we conducted on CAD reveals that the existing approaches utilizing deep learning address a wide range of applications, including cybersecurity, IoT, and key performance indicators (KPIs). Different application domains of CAD and benchmark datasets are discussed. Moreover, the commonly used datasets for CAD are described and discussed from different application scenarios. Finally, the limitations and drawbacks of the different trends used to solve the problem of CAD are outlined, and the challenges and future research directions are discussed.