Comprehensive Resource Collection for Foundation Models in Anomaly Detection: When Large Models Meet Anomaly Detection

This article introduces a carefully curated resource project on foundation models for anomaly detection, covering the applications of large language models, vision-language models, graph foundation models, time-series foundation models, etc., in the field of anomaly detection. This project addresses the limitations of traditional anomaly detection methods and provides a systematic reference for researchers, engineers, and beginners, making it worth bookmarking and following.

Anomaly detection is a classic machine learning problem applied in key areas such as financial fraud and industrial fault early warning. Traditional methods require manual feature design and struggle to adapt to complex environments. Foundation models are large-scale pre-trained, general-purpose, and fine-tunable AI models (e.g., GPT, CLIP, graph neural networks), whose strong representation and generalization capabilities bring new ideas to anomaly detection.

• Large Language Models (LLM)：Used for log semantic anomaly identification, anomaly detection in text-based representations of time series, and anomaly judgment of cross-modal multi-dimensional data;
• Vision-Language Models (VLM)：Combining images and text, applied in industrial quality inspection, anomaly description in video surveillance, and auxiliary diagnosis of medical images;
• Graph Foundation Models：Capture structural anomalies in graph structures (social networks, financial networks);
• Time-Series Foundation Models：Pre-train general time-series representations to adapt to diverse time-series anomaly detection tasks.

This project, in the form of an open-source Awesome List, collects papers and resources on the application of foundation models to anomaly detection. It organizes literature in a structured way to facilitate a quick understanding of the entire field, covering various types of foundation models and demonstrating the richness of the cross-disciplinary field.

Current cross-research trends include: multi-modal fusion, zero-shot/few-shot learning (reducing annotation dependency), interpretability (LLMs generating anomaly explanations), and real-time performance (efficient deployment). There are still open issues to explore in these directions.

• Researchers：Quickly enter the field and understand the technical context and trends;
• Engineers：Discover solutions suitable for scenarios and avoid reinventing the wheel;
• Beginners：Establish a learning roadmap for systematic understanding. It is recommended to bookmark and follow this resource list continuously.

This project establishes a knowledge hub for cross-research between foundation models and anomaly detection, demonstrating that open-source collaboration accelerates knowledge dissemination and technological progress. With the development of foundation models and the growth of application demands, more innovative results will emerge in this field, which is worth continuing to follow.