CRoSS: Diffusion Model Makes Controllable, Robust and Secure Image Steganography

Current image steganography techniques are mainly focused on cover-based methods, which commonly have the risk of leaking secret images and poor robustness against degraded container images. Inspired by recent developments in diffusion models, we discovered that two properties of diffusion models, the ability to achieve translation between two images without training, and robustness to noisy data, can be used to improve security and natural robustness in image steganography tasks. For the choice of diffusion model, we selected Stable Diffusion, a type of conditional diffusion model, and fully utilized the latest tools from open-source communities, such as LoRAs and ControlNets, to improve the controllability and diversity of container images. In summary, we propose a novel image steganography framework, named Controllable, Robust and Secure Image Steganography (CRoSS), which has significant advantages in controllability, robustness, and security compared to cover-based image steganography methods. These benefits are obtained without additional training. To our knowledge, this is the first work to introduce diffusion models to the field of image steganography. In the experimental section, we conducted detailed experiments to demonstrate the advantages of our proposed CRoSS framework in controllability, robustness, and security.

Paper

References (78)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 1iwK4/10 · confidence 4/52023-06-14

Summary

This paper presents CRoSS, a novel image steganography framework leveraging text-driven diffusion models. It offers improved security, robustness, and controllability compared to cover-based methods.

Strengths

CRoSS is the first work to introduce diffusion models to image steganography and achieves these benefits without requiring additional training.

Weaknesses

- Equation 3 should be written in detail. Is it the same as those in Section H of appendix in guided diffusion [A]? - The collected dataset only contains 260 images, rendering it not convincing in the conclusions from the presented quantitative experiments. Besides, the secret and stego image pairs much resemble either in appearance or in concept, e.g., cat versus tiger, or man versus woman. As a result, there is a restriction of secret-stego correlation that does not exist in the compared previous works. - Stable diffusion [B] utilizes VAE to reconstruct images from the latent space, leading to inherent errors. As a result, the performance of CRoSS is constrained by this factor, resulting in significant errors. In hiding tasks, the extraction accuracy is a crucial metric. But the PSNR between the revealed images and the corresponding secret images is merely 23.79dB, or even less when there is distortion, suggesting that the method can be less applicable in the real world. - For each instance of covert transmission using the proposed method, the recipient must know the exact prompt or condition, e.g., word, sketch, or even parameters, used during hiding, which might be costy and inpractical. - As stated in [C], setting the guidance scale to 1 (used in the classifier-free control algorithm) may compromise the image editability. But in this paper, it is indeed set to 1, which could potentially limit the range of secret image selection. [A] Dhariwal P, Nichol A. Diffusion models beat gans on image synthesis[J]. Advances in Neural Information Processing Systems, 2021, 34: 8780-8794. [B] Rombach R, Blattmann A, Lorenz D, et al. High-resolution image synthesis with latent diffusion models[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 10684-10695. [C] Mokady R, Hertz A, Aberman K, et al. Null-text inversion for editing real images using guided diffusion models[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 6038-6047.

Questions

- Please conduct large-scale quantitative experiments, author may refer to the the settings in InstructPix2Pix [A]. - Please conduct experiments using prompts that show low relation with the secret images. - More distortions shall be considered to verify general robustness. [A] Brooks T, Holynski A, Efros A A. Instructpix2pix: Learning to follow image editing instructions[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 18392-18402.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

NA

Authorsrebuttal2023-08-18

Looking forward to discussions with the Reviewer 1iwK

Dear Reviewer 1iwK, We appreciate your previous review time and constructive comments. In response to your concerns, we have provided explanations, clarifications, and additional experimental results in the supplementary rebuttal PDF file. We would like to know if your concerns have been adequately addressed. If you have further questions or comments, we would be more than willing to address and discuss them. Thank you for your efforts! Best regards, Authors

Reviewer g14p3/10 · confidence 5/52023-06-26

Summary

This paper addresses coverless image steganography by taking the prompt as the guidance to generate stego images using Stable Diffusion. It shows better controllability with language-driven model, better robustness and security with stronger generation power of diffusion probabilistic model. Experimental results show better performance than SOTAs.

Strengths

+) The first coverless stegnography that uses pretrained diffusion model for image steganography; +) Prompt as control of steganography

Weaknesses

-) Novelty is somewhat limited. It is obvious that using fixed sampling could always generate the same image using a pretrained diffusion model. I'd like to see contributions in terms of network structures and insights that drives coverless steganography in general. -) Not clear about how the public and private keys are used in the model. -) Missing experiments on popular datasets such as BOSSbase and BOWS2. -) Another major problem is RIIS performance better than CRoSS in some cases. Please explain them in details. It is not proper that RIIS is trained on degraded data compared to CRoSS, as both can be set under the same setting.

Questions

+) In Figure 8, what does "Shoot" mean? +) Typo error: In Algorithm1, L2, hided -> hidden +) Does CRoSS have any limitations?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

1 poor

Contribution

1 poor

Limitations

Not mentions. I'd like to hear about any potential limitations.

Reviewer Nci37/10 · confidence 4/52023-07-06

Summary

This paper introduces diffusion models to the field of image steganography. It argues the significant advantages in controllability, robustness, and security compared to cover-based image steganography methods. It utilized the power of Stable Diffusion to translate between two images without training. This paper collects a benchmark and shows experiments on it.

Strengths

1. This paper proposes a novel idea to utilize the diffusion Model to do image steganography. I think this is reasonable because pretrained diffusion models can invert an image to noise or translate between images without training. 2. The presentation is clear such as Fig.2 and Fig.3. 3. The qualitative results seem good as shown in Fig. 6. 4. This paper collects a benchmark containing 260 images, which may be useful for the community. 5. Analysis is in detail, including discussion from Sec 4.2 to Sec 4.4.

Weaknesses

1. According to null-text inversion [1], DDIM is not enough to invert a real image perfectly. In this case, how to deal with the artifacts in your method? I believe low-level artifacts are fatal for image steganography. 2. Most secrets and containers share a similar background. Does it mean we need a good image editing tool when using this method? 3. It is difficult to generate high-quality human faces with Stable Diffusion. Inverting human faces is also not stable. Why do the qualitative results in Fig. 6 seem good? [1] Null-text Inversion for Editing Real Images using Guided Diffusion Models

Questions

1. Why Secret and Container are required to share a similar background?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Please refer to the "weakness".

Authorsrebuttal2023-08-18

Looking forward to discussions with the Reviewer Nci3

Dear Reviewer Nci3, We appreciate the time you dedicated to reviewing our work and your recognition of our work. Regarding the concerns you raised, we have provided explanations in our responses. We would like to ensure that your concerns have been adequately addressed. If there are any aspects of our work that remain unclear to you, please don't hesitate to let us know. Thank you for your dedication! Best regards, Authors

Reviewer mtD26/10 · confidence 3/52023-08-02

Summary

In this work, the authors introduce an image steganography framework (CRoSS) that leverages the properties of diffusion models to enhance the security, controllability, and robustness of the steganography process. The authors show how the diffusion model can integrate with image steganography to achieve these goals without additional training. The authors claim this to be the first work to introduce diffusion models to the field of image steganography. The effectiveness of the proposed CRoSS framework was validated with different experiments where the authors demonstrated the advantages over existing methods.

Strengths

- This paper proposes to apply diffusion models to the field of image steganography, creatively combining existing ideas to overcome the limitations of traditional methods. - The paper is clear, and the authors clearly define their goals, explain their methodology, and discuss their results. - Overall, the paper was easily understandable and easy to follow. Its structure is well-organized, and the proposed method is presented clearly and makes sense. - To validate the effectiveness of the method, the authors conducted different experiments showing the advantages of the proposed approach in terms of security, controllability, and robustness. - In general, the paper is very well written.

Weaknesses

- Although the authors considered validating their method against distortion attacks, Gaussian noise distortion is not the appropriate approach to evaluate its robustness. As authors may know, Gaussian noise is already used in the diffusion process due to its mathematical properties. The Gaussian distribution is symmetric and has the property that the sum of multiple Gaussian random variables is also Gaussian. This makes it mathematically convenient, especially in diffusion models where noise is added iteratively. Therefore, adding more Gaussian noise to the container image will not affect the reveal process too much. I would like to see if the method is still robust to another type of noise but Gaussian. - In my opinion, the robustness validation made in the paper is not enough to conclude that the method is robust. For example, did the authors consider the scenario where an attacker performs not only global distortions (like JPEG compression, adding noise, etc.) but local distortions, e.g., blurring a patch on the image?

Questions

- Will the proposed method recover the secret image well when an attacker applies another type of noise? Can this negatively impact the performance of the method? - Is the method robust to local distortions?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors acknowledged the limitations of their work, including the gap in pixel-wise objective fidelity metrics, the trade-off that sacrifices the editing capability to ensure the zero-shot invertibility of image translation, and the limitation of hiding only one secret image within a single container image. However, they could provide more insights into potential solutions or future research directions to address these limitations. They did not explicitly discuss potential negative societal impacts.

Reviewer g14p2023-08-11

Response

I appreciate the responses from authors. However, I have concerns regarding: 1. Steganalysis. A well steganography method is capable of defending various steganalysis methods. The popular SRNet [1] and SigstegNet [2] are needed to validate performances of the steganography method. I believe SiastegNet is an improved version of SID. 2. I believe CRoSS and RIIS can share the same setting when both of them are trained on the same setting. Otherwise you could not compare the two. According to the numeric results, I'm afraid it's no better than RIIS. [1] Mehdi Boroumand, et al, Deep Residual Network for Steganalysis of Digital Images, http://www.ws.binghamton.edu/fridrich/research/SRNet.pdf [2] Weike You et al, A Siamese CNN for Image Steganalysis, TIFS

Authorsrebuttal2023-08-12

Response to Reviewer g14p's Concerns

Thank you for your prompt response. Below is our response for your concerns: > ***Experiments involving various steganalysis methods*** - Firstly, our paper $\textbf{\textcolor{red}{already}}$ includes experiments involving various steganalysis methods, as shown in $\textbf{\textcolor{red}{Figure 5 and Table 1}}$ of the main paper. These methods include **SID** [1], **StegExpose** [2], **XuNet** [3], **YedroudjNet** [4], and **KeNet** [5]. $\textbf{\textcolor{red}{It's worth noting that KeNet [5] is the same as the method SigStegNet you mentioned.}}$ The authors of [5] chose the name **"KeNet"** for **SigStegNet**, and you can find more details about this naming choice in their GitHub link [6]. - Secondly, we have included experiments involving detection accuracy of **SRNet** [7] as you mentioned. The closer the detection accuracy is to 50%, the better the security. The results of these experiments are as follows and demonstrate the superiority of CRoSS over other methods in terms of security: | #Leaked Samples | 60 | 120 | 180 | | ----- | ----- | ----- | ----- | | ISN | 57.5% | 67% | 66.75% | | Baluja | 55.75% | 65% | 71.5% | | HiNet | 58.67% | 63.5% | 68.33% | | RIIS | 54.25% | 61.67% | 59.5% | | **CRoSS** | **54%** | **59.5%** | **57.25%** | - In conclusion, we believe that the number of steganalysis methods involved in the experiments is $\textbf{\textcolor{red}{sufficient}}$ to demonstrate the security advantages of CRoSS. > ***Experiments with the same settings for both RIIS and CRoSS*** - We need to emphasize again that $\textbf{\textcolor{red}{CRoSS does not require training for specific distortion types.}}$ In order to make CRoSS and RIIS share the same setting, we compared the robustness differences between RIIS and CRoSS, where RIIS **was not trained under any distortion types.** The PSNR results are shown below, and CRoSS outperforms RIIS (**"GN"** refers to Gaussian noise): | Methods | GN ($\sigma=10$) | GN ($\sigma=20$) | GN ($\sigma=30$) | JPEG ($Q=20$) | JPEG ($Q=40$) | JPEG ($Q=80$) | | --- | --- | --- | --- | --- | --- | --- | | RIIS | 13.78 | 12.73 | 11.09 | 7.48 | 10.05 | 13.82 | | **CRoSS** | **21.89** | **20.19** | **18.77** | **21.74** | **22.74** | **23.51** | - Furthermore, based on the experimental results we have already provided, CRoSS $\textbf{\textcolor{red}{undoubtedly outperforms}}$ RIIS in terms of robustness and security. These results include: - Most results in $\textbf{\textcolor{red}{Figure 5 and Table 1}}$ of the main paper (security experiments). - $\textbf{\textcolor{red}{Figure 8}}$ of the main paper (subjective results of robustness experiments). - Most results in $\textbf{\textcolor{red}{Table 2}}$ of the main paper (objective results of robustness experiments). - $\textbf{\textcolor{red}{Figure B.4}}$ in the supplementary material (more subjective results of robustness experiments). - $\textbf{\textcolor{red}{Figure 1, Figure 3 and Table 1}}$ in supplementary rebuttal PDF file (more subjective and objective results of robustness experiments). > ***References*** [1] Steganalyzing images of arbitrary size with CNNs. Electronic Imaging 2018. [2] Stegexpose-A tool for detecting LSB steganography. ArXiv 2014. [3] Structural design of convolutional neural networks for steganalysis. IEEE Signal Processing Letters 2016. [4] Yedroudj-net: An efficient CNN for spatial steganalysis. ICASSP 2018. [5] A Siamese CNN for image steganalysis. TIFS 2020. [6] https://github.com/SiaStg/SiaStegNet#quickstart [7] Deep residual network for steganalysis of digital images. TIFS 2018.

Reviewer g14p2023-08-18

Response

I read the rebuttal. Authors have address my first question well, however, I still maintain my concern that the comparison between RIIS and CROSS is not fair. Besides, I noticed that Reviewer g14p has mentioned that extraction accuracy for hiding task is quite important. The PSNR of the revealed image hardly reaches 40dB, which is not good. Therefore, I maintain my initial rating.

Authorsrebuttal2023-08-19

Response to Reviewer g14p's Concerns

Thank you for your response. We need to emphasize again that: - Evaluating an image steganography algorithm requires a $\textcolor{red}{\textbf{comprehensive assessment of fidelity, security, and robustness.}}$ Dismissing CRoSS **solely** based on fidelity **under ideal conditions** is one-sided and unreasonable. - Under ideal conditions, RIIS, as a method that **requires training**, indeed exhibits better fidelity than CRoSS, which **doesn't require training**. However, in practical scenarios, distortions inevitably appear on container images, and image steganography algorithms face the risk of detection. At this point, the performance in terms of **security and robustness** becomes particularly crucial. CRoSS holds a distinct advantage in these aspects, which is a valuable insight we gained by introducing diffusion models into image steganography. $\textcolor{red}{\textbf{Comparing CRoSS and RIIS in real-world application scenarios is reasonable}}$, and we fail to comprehend why you consider this comparison unfair.

Authorsrebuttal2023-08-18

Looking forward to further discussions with the Reviewer g14p

Dear Reviewer g14p, We appreciate your previous review and the prompt response. We have responded to your latest concerns and supplemented the experiments on steganalysis methods and the comparison between RIIS and CRoSS. We would like to further discuss whether your concerns have been addressed. If there are any aspects of our work that are still unclear to you, please let us know. Thank you for your continued engagement! Best regards, Authors

Reviewer mtD22023-08-11

I thank the authors for their detailed response. The provided rebuttal addressed my questions. I will update and maintain my score toward acceptance.

Authorsrebuttal2023-08-12

Thank Reviewer mtD2 for recognizing our work

Dear Reviewer mtD2, Thank you for engaging in our discussion and recognizing that our responses have effectively addressed all the concerns you raised. We greatly appreciate your constructive comments, which have contributed to the improvement and solidity of our work! Best regards, Authors

Reviewer 1iwK2023-08-19

Response

I have read the rebuttal and some concerns such as Weakness #2, have been addressed. However, I still maintain some issues as follows: + Weakness #1: Actually, Appendix H in Guided Diffusion can be viewed as a detailed representation of the DDIM-sampled ODE equation. And diffusion models can be applied with diverse ODEs, such as EDM[1]. Therefore, a detailed description is warranted. + Weakness #3#5: The objective of a steganographic system is to accurately convey secret messages. Accuracy on extraction is an indispensable metric. Moreover, cover and stego images often share some visual elements like backgrounds, which can compromise security to a certain extent. Setting the guidance scale to 1, in my opinion, restricts the breadth of real-world application scenarios. + Weakness #4: The authors mentioned that during the reveal process, it's possible to guess the prompt key. If feasible, it would be helpful to showcase relevant examples and demonstrate the extraction results. Thanks for the authors' response. Looking forward to further discussion. **Reference**: [1] Elucidating the design space of diffusion-based generative models. NIPS 2022.

Authorsrebuttal2023-08-19

Response to Reviewer 1iwK's Concerns (part 1)

Thank you for your timely response and active engagement in the discussion! Below is our latest response: > ***Weakness #1: The details of Equation 3 and its relationship with appendix H of guided diffusion*** We have carefully read Appendix H of the guided diffusion paper [1] (titled "Conditional Diffusion Process") and would like to provide some clarifications regarding your comments: - Appendix H of [1] provides an overview of the theory behind the **conditional diffusion model**. The conditional diffusion model can be represented by $q(x_{t}|x_{t+1}, y)$. Using Bayes' theorem, it can be further broken into two components: $q(x_{t}|x_{t+1})$ representing the unconditional diffusion model and $q(y|x_{t})$ representing conditional guidance. For more detailed insights, please refer to Equation 55~61 in Appendix H of [1] along with the subsequent analysis. It's important to emphasize that the theory of the conditional diffusion model discussed in Appendix H of [1] is $\textcolor{red}{\textbf{independent of specific sampling algorithms}}$, whether it's DDIM or DDPM, ODE or SDE. - Our Equation 3 describes the general DDIM sampling process, which is applicable to both conditional and unconditional diffusion models. In summary, there is a **noticeable distinction** between Equation 3 and Appendix H of [1]. However, we also find your suggestion to provide more detailed description for Equation 3 constructive. We will incorporate your suggestion into the revised version. > ***Weakness #3&#5: Issues of accuracy on extraction and the setting of "guidance scale=1"*** - Regarding the issue of fidelity, or more specifically, the **accuracy on extraction**, our clarification is as follows: - Firstly, we agree with your perspective that fidelity is a crucial aspect of image steganography algorithms. However, achieving higher PSNR solely **under ideal conditions** falls short of practical significance. In real-world scenarios, container images often carry complex real-world distortions. Investigating fidelity under such conditions is more practical and significant. However, $\textcolor{red}{\textbf{this point has been overlooked in previous works}}$, which is precisely where CRoSS holds a unique advantage. CRoSS demonstrates better fidelity under various distortion conditions, showcasing enhanced robustness. - Additionally, we believe that pixel-wise evaluation metrics like PSNR can be misleading. A significant amount of research has concentrated on boosting PSNR **under ideal conditions**, disregarding real-world application needs. While our CRoSS may not achieve the same level of **objective fidelity** as compared to previous methods, its **subjective fidelity** in terms of visual perception is indeed satisfactory. - Lastly, $\textcolor{red}{\textbf{there is substantial room for improvement in CRoSS's objective fidelity.}}$ There are two possible approaches: (1) Not using latent diffusion models like Stable Diffusion but employing image diffusion models such as Imagen [2]. (2) Introducing strategies that require training. It's worth noting that CRoSS currently operates **without requiring training**. However, by maintaining the overall framework and simultaneously incorporating dedicated training for image steganography tasks, it is possible to mitigate its objective fidelity disadvantage while preserving its advantages in robustness and security. This would be a valuable topic for the future research. - Regarding the issue of background similarity and the setting of **"guidance scale=1"**, our clarification is as follows: - Currently, in order to ensure subjective fidelity, we have adopted the Stable Diffusion with a guidance scale of 1. This has indeed led to certain issues such as background similarity. What we want to emphasize here is that these issues $\textcolor{red}{\textbf{do not stem from the CRoSS framework itself}}$, but rather from the specific choice of the diffusion model used. - Because the editing capability of Stable Diffusion with a guidance scale of 1 is limited, we adopted experimental settings such as using similar backgrounds. However, the generative capability of diffusion models is **continuously evolving**. For instance, recent advancements like the upgraded version of Stable Diffusion (SDXL) are further pushing the boundaries of diffusion model capabilities. In the future, we hope to leverage **more powerful** classifier-free diffusion models with a guidance scale of 1, or even non-classifier-free diffusion models, to achieve **stronger editing capabilities**. - This way, we can enhance the CRoSS framework in a **plug-and-play manner**. CRoSS, as the first effort to integrate diffusion models into image steganography, introduces a new framework, showcasing pioneering value. > ***References*** [1] Diffusion Models Beat Gans on Image Synthesis. NeurIPS 2021. [2] Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. NeurIPS 2022.

Authorsrebuttal2023-08-19

Response to Reviewer 1iwK's Concerns (part 2)

> ***Weakness #4: Relevant examples when the receiver guesses the prompt*** - In the main paper, in **Figure 4** on the far left (Scenario #1), we illustrate potential candidate revealed results that the receiver might generate when trying to guess the prompt. As can be observed, for the receiver, it is not feasible to deduce the correct answer through random guessing of the prompt, as each candidate revealed image appears authentic. - We will incorporate your instructive suggestion and provide more examples in the revised version. We hope that our responses have addressed all of your concerns. We sincerely appreciate your continued engagement!

Authorsrebuttal2023-08-21

Looking forward to further discussions with the Reviewer 1iwK

Dear Reviewer 1iwK, Thank you for your continued engagement in the discussion! We apologize for any inconvenience caused by reaching out again. However, today marks the last day of the discussion phase, and we have provided detailed analysis and clarifications in response to your latest concerns. We would like to further discuss with you to ensure that your concerns are addressed. Looking forward to our continued discussion! Best regards, Authors

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC