Anomaly detection using Diffusion-based methods

TL;DR

The paper addresses anomaly detection in challenging, high-dimensional data by evaluating diffusion-based models, specifically DDPMs and Diffusion Transformers, using reconstruction objectives. It demonstrates that these models achieve superior adaptability, scalability, and robustness over traditional methods like Isolation Forest, One-Class SVM, and COPOD on both compact and high-resolution datasets. Reconstruction error emerges as a strong signal for detecting anomalies, with diffusion-based architectures effectively handling noisy and adversarial scenarios and scaling to large datasets such as Mini-ImageNet. The work suggests diffusion-based anomaly detection is a powerful, scalable baseline and points to future work in encoder-decoder design and multi-modal data integration to broaden applicability.

Abstract

This paper explores the utility of diffusion-based models for anomaly detection, focusing on their efficacy in identifying deviations in both compact and high-resolution datasets. Diffusion-based architectures, including Denoising Diffusion Probabilistic Models (DDPMs) and Diffusion Transformers (DiTs), are evaluated for their performance using reconstruction objectives. By leveraging the strengths of these models, this study benchmarks their performance against traditional anomaly detection methods such as Isolation Forests, One-Class SVMs, and COPOD. The results demonstrate the superior adaptability, scalability, and robustness of diffusion-based methods in handling complex real-world anomaly detection tasks. Key findings highlight the role of reconstruction error in enhancing detection accuracy and underscore the scalability of these models to high-dimensional datasets. Future directions include optimizing encoder-decoder architectures and exploring multi-modal datasets to further advance diffusion-based anomaly detection.

Figures (1)

• Figure 1: AUC Scores for Anomaly Detection Methods on the High-resolution dataset (Scalability Test). The superior performance of our method shows that Diffusion Transformers are more scalable compared to conventional methods.