Survey Maps Multi-Modal Anomaly Detection Landscape.

This paper presents a comprehensive survey of Multi-Modal Anomaly Detection (MMAD), a critical field for identifying rare, abnormal events from diverse data sources in applications like industrial inspection and cybersecurity. The existing literature on MMAD is often fragmented across different domains and modality combinations, and previous surveys typically group methods by architecture rather than by how anomalies are fundamentally defined and separated in multi-modal contexts. The authors formalize the MMAD problem, identifying five intrinsic characteristics that underpin its core challenges. They then organize prior work into two complementary paradigms: "normality-assumption methods" and "anomaly-assumption methods." The former focuses on modeling regularity through representation learning, cross-modal alignment, and knowledge enhancement, while the latter aims to sharpen decision boundaries by injecting coarse-grained, structural, and semantic anomalies. The survey also investigates the transformative impact of foundation models on MMAD, noting their contributions through scalable pretraining, flexible cross-modal transfer, and emerging reasoning capabilities. Finally, it compiles representative benchmarks and evaluation protocols across various domains and outlines open problems and future research directions for developing more robust, adaptive, and interpretable MMAD systems.