MDDM: Practical Message-Driven Generative Image Steganography Based on Diffusion Models

Authors: Zihao Xu, Dawei Xu, Zihan Li, Chuan Zhang

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that MDDM offers notable advantages in terms of accuracy, controllability, practicality, and security. [...] 5. Experiments Our experiments are conducted based on publicly available pre-trained diffusion models.
Researcher Affiliation Academia 1Changchun University 2Beijing Institute of Technology. Correspondence to: Chuan Zhang <EMAIL>.
Pseudocode Yes Algorithm 1 Generate initial noise Input: The noise size (1, c, h, w), secret message m with length l, Cardan grille CG, truncation threshold k Output: Initial noise x T Function F1(n {0, 1}) list N(0, I) if n = 0 then y = Random Choice{u | u < k, u list} else y = Random Choice{u | u > k, u list} end if Return y Function F2() list N(0, I) y = Random Choice{u | k u k, u list} Return y Init noise x T for i < l do x T [CG[i]] = F1(m[i]) end for for j < 1 c h w do if j / CG then x T [j] = F2() end if end for
Open Source Code No The paper does not provide an explicit statement or a direct link to the source code for the MDDM methodology. It mentions using publicly available pre-trained diffusion models and code from other methods like Pulsar, but not its own implementation.
Open Datasets Yes We compare MDDM with two GAN-based generative image steganography methods, Stega Style GAN (Su et al., 2024) and CIS-Net (You et al., 2022), on the Celeb A dataset (Liu et al., 2015). [...] For unconditional image generation, we use pre-trained diffusion models on the Celeb A-HQ (face images) (Karras et al., 2017), LSUN-Bedroom, and LSUN-Cat (Yu et al., 2015) datasets.
Dataset Splits No The paper refers to standard datasets like Celeb A, Celeb A-HQ, LSUN-Bedroom, and LSUN-Cat, but does not explicitly provide specific training/test/validation split percentages, sample counts, or detailed splitting methodologies used for its own experiments beyond mentioning the datasets themselves. Appendix D.3 details experimental setup for steganographic capability but also does not specify dataset splits.
Hardware Specification Yes All experiments are conducted on NVIDIA RTX 3090 GPUs.
Software Dependencies No MDDM requires no additional training and instead relies on multiple pre-trained modules. Specifically, it leverages either latent diffusion models (e.g., Stable Diffusion) or pixel-space diffusion models (e.g., DDPM), together with a DDIM scheduler and inversion methods (such as DDIM inversion and EDICT), all of which are readily available. The paper refers to model versions (e.g., Stable Diffusion v1.5, SD-V1.4, SD-V2.1) but does not list specific software dependencies like programming languages, frameworks, or libraries with their version numbers.
Experiment Setup Yes The number of steps for both sampling and inversion is set to 50. [...] in subsequent experiments based on Stable Diffusion, we use the guidance scale of 7.5 by default to conditionally generate stego images. [...] The prompts for the Face, Bedroom, and Cat categories are Portrait photo, best quality, masterpiece, ultra detailed, UHD 4K, photographic, 1girl, face, looking at viewer, color photo, natural colors , A photo of the bedroom , and A cat , respectively. [...] The diffusion model employed is Stable Diffusion v2.1, with image resolution set to 512 512 and an empty prompt.