Multivariate Time Series Anomaly Detection using DiffGAN Model

TL;DR

This work tackles anomaly detection in multivariate time series (MTS-AD) and the sensitivity of diffusion-step choices in partial-diffusion models. It introduces DiffGAN, a GAN-augmented denoiser that replaces the forward diffusion with a generator while the discriminator predicts the necessary denoising steps, and retains a diffusion-based denoiser for reconstruction. The approach jointly learns noise generation, step guidance, and high-quality reconstruction, achieving superior F1 scores versus state-of-the-art reconstruction models across multiple MTS datasets and diffusion-step settings. Overall, DiffGAN provides a flexible, efficient framework for robust MTS-AD with improved reconstruction and detection performance.

Abstract

In recent years, some researchers have applied diffusion models to multivariate time series anomaly detection. The partial diffusion strategy, which depends on the diffusion steps, is commonly used for anomaly detection in these models. However, different diffusion steps have an impact on the reconstruction of the original data, thereby impacting the effectiveness of anomaly detection. To address this issue, we propose a novel method named DiffGAN, which adds a generative adversarial network component to the denoiser of diffusion model. This addition allows for the simultaneous generation of noisy data and prediction of diffusion steps. Compared to multiple state-of-the-art reconstruction models, experimental results demonstrate that DiffGAN achieves superior performance in anomaly detection.

Figures (3)

• Figure 1: Architecture of DiffGAN.
• Figure 2: The reconstruction and anomaly detection outcomes of TadGAN and DiffGAN on a segment of the Global dataset.
• Figure 3: F1 score varies with the number of diffusion steps using diffusion model on the Global dataset.