A Coverless Steganography of Face Privacy Protection with Diffusion Models

A Coverless Steganography of Face Privacy Protection with Diffusion Models

연구 분야: Strategies

학회: International Conference on Pattern Recognition

As a highly recognizable biometric feature, human face has become the first choice for identity verification. With the application of face in various important fields of society, the serious threat caused by face image information leakage has become prominent, and its privacy and security protection is becoming more and more important. Applying steganography to face images can not only effectively protect personal privacy, but also realize the secure transmission and sharing of sensitive information. Therefore, we propose a face privacy-preserving coverless steganography framework based on diffusion models. Firstly, the facial features are extracted and the feature masks are generated. Then, the DDIM sampling is used to generate the coverless stego image by combining the conditional diffusion model with the text secret key by using the generation ability of diffusion model. DDIM Inversion is used to recover the secret image with high quality. We conduct extensive experiments on CelebA-HQ and FFHQ public face datasets. Compared with the existing methods, the stego images generated and recovered by our method have higher quality and can better resist steganalysis. Our method also achieves significant advantages in terms of robustness and security, maintaining sharper recovery effects under Gaussian noise, JPEG compression, and real-world transmission. In addition, we can combine custom masks to achieve controllable local steganography, which has stronger controllability and flexibility. The proposed method can achieve a good unity of security, controllability and robustness, and is superior to the traditional steganography methods without any additional training.