A reconstruction-based approach to multivariate time series anomaly detection that replaces the fixed partial-diffusion schedule of diffusion-model detectors with a generative adversarial mechanism. A generator produces the noised version of an input window in place of the forward diffusion process, while the discriminator doubles as a controller that predicts how many denoising steps the diffusion denoiser should run for that input. Anomalies are scored by reconstruction error between the input and its adaptively denoised reconstruction. Covers why the choice of diffusion step count governs reconstruction fidelity and detection quality, and how coupling GAN and diffusion components sidesteps hand-tuning it.
DiffGAN is a reconstruction-based method for multivariate time series anomaly detection that grafts a generative adversarial network onto the denoiser of a diffusion model. Standard partial-diffusion…