Foundation Model-Energized Anomaly Detection and Outlier Detection: A Survey

Anomaly detection (AD), the task of identifying samples that significantly deviate from normative data patterns, is a critical challenge in data analysis and decision-making. It has widespread applications in high-stakes fields such as financial fraud detection, medical diagnosis, and cybersecurity intrusion detection. While AD research has long been active in machine learning, the recent rise of foundation models (FMs) has spurred a surge of new, FM-based techniques. This proliferation of methods provides a diverse toolkit but also creates confusion for learners and practitioners. Consequently, a systematic review is urgently needed to synthesize the existing literature, clarify the current research landscape, and identify persistent challenges. This survey addresses three key aspects of FM-energized anomaly detection: (i) the methodological approaches for leveraging FMs in AD; (ii) the comparative advantages and limitations of FM-based methods against traditional techniques; and (iii) the prevailing trends, thematic preferences, and unresolved challenges in the field. Based on this tripartite analysis, we identify promising yet underexplored research pathways. We anticipate that this survey will not only elucidate current developments and trajectories but also catalyze further innovation within the AD community.