An image steganography algorithm using selective timestep embedding and diffusion model

doi:10.1016/j.eswa.2026.131313

Return

An image steganography algorithm using selective timestep embedding and diffusion model

2026-01-21

PRE

Jiajun Han

Guodong Ye *

Sirui Han *

DOI：10.1016/j.eswa.2026.131313

Original

Original request for help

Save

Abstract

En 中文

Generative steganography has emerged as a promising scheme for enhancing the security of covert communication. However, existing diffusion-model-based approaches primarily utilize the model as a static generator or embed secret data exclusively at the initial noise input, failing to exploit the dynamics intrinsic to the denoising process. To address this limitation, an image steganography algorithm is presented using selective timestep embedding and diffusion model, named ImSA. Unlike conventional methods, ImSA introduces a plaintext related mechanism to select the optimal timestep, embedding secret information into the frequency domain of the intermediate latent variables. Specifically, the plain image is encrypted via a hyperchaotic system into noise-like sequence and subsequently embedded into the high-frequency wavelet subband at the selected timestep. Furthermore, the diffusion model is retrained to adapt to the distribution of latent variables containing the secret data. Compared with current generative steganography methods, the contributions of ImSA are: (1) The imperceptibility and security can be improved by dynamically selecting the specific timestep embedding; (2) The high-capacity information embedding and high-quality images generation can be achieved by frequency-domain embedding; (3) Minimizes disruption to the generative process by converting the plain image into a noise-like cipher. Moreover, experimental results demonstrate a 2.0 bpp embedding capacity with high generation quality, robust extraction quality (PSNR with 38 dB) and strong security performance (Probability of Error with 0.51).