The widespread use of deep neural networks in various algorithms has raised concerns regarding their security vulnerabilities. Recently, diffusion models have been used to purify adversarial noise from images because of their powerful generative abilities. However, traditional diffusion models rely on a single Gaussian noise model during the forward process, which may be insufficient for capturing the complex distributions of data, particularly in the presence of adversarial noise. This paper proposes a Gaussian Mixture Model-based Diffusion Model for Purification (GMDMP). The approach introduces a Gaussian Mixture Model (GMM) into the diffusion process of a Denoised Diffusion Probabilistic Model (DDPM), leveraging the GMM’s more flexible distributional properties to capture complex noise disturbances more accurately. Extensive comparative experiments on both CIFAR-10 and ImageNet datasets demonstrate that the proposed approach outperforms the state-of-the-art baselines. On CIFAR-10 and ImageNet, it achieves 79.88% and 56.45% average robust accuracy, which are 8.59% and 12.56% higher, respectively.