Enhancing Transferability via Spectral Alignment and Grad-CAM Guided Latent Diffusion Models

With the rapid development of diffusion models, adversarial examples generated using these models exhibit high quality and have become an emerging attack strategy that is widely studied. However, images generated by diffusion models still exhibit spectral differences from the original, leaving room for improved transferability in black-box scenarios. This paper proposes a method called Enhancing Transferability via Spectral-Aligned and Grad-CAM Guided Latent Diffusion (SAGC-LDiff) to improve the success rate of adversarial attacks in cross-model scenarios. By introducing the Control-Variational Autoencoder (VAE) module and utilizing DFT and DCT to extract high-frequency and low-frequency features of the image, this method reduces the spectral differences between the original and generated images, thereby minimizing the artifacts introduced by VAE and improving the success rate of cross-model transfer attacks. In the latent space denoising stage, the salient region latent mask generated by Grad-CAM guides the generation of adversarial examples, making these examples more natural and realistic. Experimental results show that, for the ILSVRC2012 dataset, SAGC-LDiff significantly outperforms existing methods on Convolutional Neural Networks and Vision Transformer models.