A Stability-Aware GAN-Diffusion Framework for Mitigating Adversarial Attacks

With the development of generative models, face generation has gradually become a research hotspot in this field because face data is needed to train the generating model. However, the data generated by the current generative model, whether the generative adversarial network or the diffusion model, often exposes the original data of the model training due to different degrees of attack. This resulted in the leaking of private facial data originally used to train the model. This challenges the security and stability of the model. Therefore, the GAN-Diffusion framework (GDF) based on the generative adversarial network (GAN) and the Latent Diffusion model (LDM) for facial privacy protection is proposed. The framework prevents the leakage of the original data by making it more difficult to attack the diffusion model by linking the GAN and the Latent Diffusion Model through a special discriminator and directing the framework to output faces that are less similar to the original data. Our proposed GDF framework is trained on 500 images on VGGFace2, and the experimental results show that only 57 images are destroyed, indicating that the method improves the privacy protection effect and maintains a high image generation quality.