Diffusion models generate diverse synthetic images to test neural network robustness, significantly reducing accuracy.