Self-supervised enhanced denoising diffusion for anomaly detection

Abstract

Generative models have significantly enhanced anomaly detection through their powerful ability to model data. However, many existing generative model-based anomaly detection methods prioritize refining the generative process while overlooking the importance of acquiring discriminative data representations, which is suboptimal. Furthermore, the intricate architectures of these methods contribute to poor model convergence, resulting in coarse data reconstructions and diminished performance. To address the above problems, we propose self-supervised enhanced denoising diffusion for anomaly detection (SDAD) to detect anomalies effectively. Specifically, SDAD acquires discriminative data representations through an auxiliary learning module with two pretext tasks, facilitating the distinction between normal data and abnormal data. Subsequently, a denoising diffusion module is used to accurately learn the normal data distribution, serving as a benchmark for anomaly detection. Extensive experiments on ten real-world datasets demonstrate the remarkable advantages of SDAD over nine state-of-the-art anomaly detection methods, as evidenced by significant improvements across various evaluation metrics.