This invention discloses a generative
image steganography method based on a conditional
diffusion model, SD-Stego. The
algorithm is based on the Guided-
Diffusion model, using controllable conditional text as input. During the inverse
generation process of the latent
diffusion model, an embedding and encoding strategy is designed. DenseNet deep connections and SEBlock channel self-attention mechanisms are introduced to fuse the intermediate generation in the latent space with the message matrix, achieving efficient embedding of steganographic information. Then, a frozen VAE decoder completes the steganographic
image generation. In the decoding stage, a multi-layer convolutional network is used to extract high-level features of the generated
image layer by layer. Batch normalization and non-linear activation functions are combined to extract embedded information, and the information is gradually mapped through flattening and fully connected
layers, finally outputting the decoded message vector. To optimize the
steganography effect and generation quality, a joint
loss function is designed, including decoding loss and
image quality loss, ensuring the stealth of the generated image and the accuracy of message decoding. This invention enables
steganography without the need for a carrier image, making it suitable for
information security and data
steganography. By utilizing denoising generation and text guidance through a conditional
diffusion model, the steganography process exhibits advantages such as high robustness, excellent
image generation quality, and
controllability of generated content. It is more suitable for complex real-world scenarios (such as
social network transmission) and has broad application prospects.