# Diffusion based Steganography Without Embedding

## Problem Statement and Motivation

Image steganography, the practice of hiding a secret image in a cover image, is an important technique for securing digital communication. Unlike cryptography-based approaches, steganography aims to avoid drawing unwanted attention from hackers by making it difficult to detect the presence of a secret message. Recent research has shown that deep learning-based approaches can improve the quality of the container image and make the changes between the cover and container image imperceptible. Nevertheless, keeping the container image undetected remains an unsolved problem. One promising approach involves using deep learning-based generative models for steganography without embedding, i.e. generating the container image without embedding the secret image directly. Recent research has shown that latent diffusion models can generate high-quality images while allowing for control over specific image features. However, the effectiveness of applying latent diffusion models to steganography without embedding remains unknown. The objective of this study is to investigate the feasibility of using latent diffusion models to generate container images for steganography without embedding. By doing so, this study aims to contribute to improving the imperceptibility and security of steganography techniques potentially benefiting the domains of information security, privacy, and communication.

## Literature Review

### Reversible Data Hiding

Reversible data hiding (RDH) is a technique that enables the embedding of additional data into a host signal (such as an image or video) in a way that can be completely recovered without any loss of the original host signal. RDH schemes are designed to be fully reversible, meaning that the original host signal can be fully restored after the hidden data is extracted.

RDH techniques are used for a variety of applications[^12][^23], including digital watermarking[^7], content authentication, and image steganography[^17][^22]. In image steganography, RDH schemes are used to embed a secret message into a cover image in such a way that the existence of the message is not detectable by an observer who does not have the appropriate decoding key.

### Steganography

Image steganography[^16] is a technique of hiding a secret message or image within a cover image, in a way that the existence of the hidden message is not detectable to unauthorized viewers. The goal is to make the stego-image (i.e., the cover image with the embedded secret message) look as similar as possible to the original cover image, so that the hidden message cannot be easily detected by anyone who doesn't have the appropriate decoding key.

This can be done by replacing the least significant bits (LSBs) of the cover image with the secret message bits, or by using more advanced techniques that take into account the statistical properties of the cover image to minimize the distortion caused by the embedding process. Recently a deep learning based approach for Image Steganography was proposed in[^17] where a hiding network encodes a secret image in feature space into a container image which is then reconstructed using a revealing network.

### Steganalysis

Steganalysis is the process of detecting the presence of hidden information within a cover image. In the context of image hiding, steganalysis techniques are used to analyze the stego-image (i.e., the cover image with the embedded secret message) and determine whether or not it contains hidden data. The goal of steganalysis is to detect the presence of hidden data even if the stego-image appears visually identical to the original cover image.

There are various steganalysis techniques used in the field of image hiding, including statistical analysis, visual inspection, and machine learning. Steganalysis is an important area of research in image hiding, as it helps to determine the security and robustness of different image hiding techniques.

Steganalysis using deep learning has emerged as a promising direction to detect and destroy hidden messages in stego-images. For instance[^15] uses siamese networks for learning to contrast between cover and stego-images,[^25] uses residual networks for steganalysis,[^10] trains an autoencoder network for destroying the hidden message encoded using a GAN based objective.

### Applications of Steganography

Some common applications of steganography in general and image steganography in particular, based on literature and research are secure communication over an insecure channel where messages can be intercepted, digital watermarking for copyright protection[^7], authentication, backdoor attacks[^1] on neural networks using steganography[^8], neural style transfer[^6].

These are just a few examples of the many applications of steganography, both in general and specifically for image steganography. As technology advances and new techniques are developed, it is likely that new applications of steganography will emerge.

### Deep Learning in Steganography

Deep learning-based steganography techniques are becoming increasingly popular in recent years. Achieving lower error rates when retrieving images, a higher payload capacity[^18][^20], robustness to distortions[^21], domain independence[^24] and imperceptibility[^11][^19] have been active directions of research in this field. Deep learning approaches beacuse of their ability to learn patterns are able to recover secret embeddings even when the changes are imperceptible to humans. In encoder-decoder architectures[^2][^4] a hiding network (encoder) is used for hiding a secret message in a container image which is then retrieved from the container image using the revealing (decoder) network. The problem has also been modeled as a 3 player game[^5][^9][^22] where in addition to encoding and decoding the images, a critic is used to guess whether the input image contains a secret or not, further improving the imperceptibility of images by guiding the stego-image distribution to be similar to the cover images.These approaches are limited by the domain that the hiding network has been trained on. Style Transfer based steganography[^3] encodes the secret image as the style in the target image which can then be extracted by a decoding network to recover the secret message. Invertible neural networks[^18][^20] offer a higher payload capacity by modeling data hiding as a reversible process using concealing and revealing networks with shared parameters. Instead of embedding the secret messeage into a container image, Steganography Without Embedding (SWE)[^19] maps a secret message to an image perceptually similar to a real image from which the secret message can then be extracted. Imperceptible adversarial perturbations[^24] used to train an image such that the network reveals the secret by picking up the perturbations as signals can generalize rather than restricting the secret message or cover images to a particular domain.

## Methodolgy (pick one)[^13][^14]
- Use LDM architecture without the noise component for better latent representation learning given the shift in research direction (still quite old)
- Direction to explore further: Use LDM generative model for generating imperceptible stegano images (retrieval part is important here, also intuition behind this?)
- Formulating the problem as SWE (Steganography without embedding) using latent stable diffusion for image generation for the texture vector (The structure vector contains the hidden message information and the texture vector is for generating diverse images using a stable diffusion model)
- IDEAS[^19] presented a structure texture disentanglement based approach that generated imperceptible stegano images rather than encoding images.
- Given the superiority of latent diffusion models at generating realistic images, why not replace the image generation module with latent diffusion based image generation.



## Work done till now

Literature Review and choosing a baseline.

## Work to be done

- Verifying baseline results,
- Research idea implementation,
- Working on improving the baseline model

## Outcome Expected

An improvement in the imperceptability of the secret message compared to the previous baselines. Research paper, atleast a preprint on arXiv ``¯\_(ツ)_/¯``.

## References

[^1]: Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, Dawn Song, 2017. Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning

[^2]: Seyed Hesam Odin Hashemi, Mohammad-Hassan Majidi, Saeed Khorashadizadeh, 2022. Color Image steganography using Deep convolutional Autoencoders based on ResNet architecture

[^3]: Donghui Hu, Yu Zhang, Cong Yu, Jian Wang, Yaofei Wang, 2022. Image Steganography based on Style Transfer

[^4]: Pin Wu, Yang Yang and Xiaoqiang Li, 2018. StegNet: Mega Image Steganography Capacity with Deep Convolutional Network

[^5]: Mehdi Yedroudj, Frédéric Comby, Marc Chaumont, 2019. Steganography using a 3-player game

[^6]: Hung-Yu Chen, I-Sheng Fang, Wei-Chen Chiu, 2018. Self-Contained Stylization via Steganography for Reverse and Serial Style Transfer

[^7]: Shehzeen Hussain, Nojan Sheybani, Paarth Neekhara, Xinqiao Zhang, Javier Duarte, Farinaz Koushanfar, 2022. FastStamp: Accelerating Neural Steganography and Digital Watermarking of Images on FPGAs

[^8]: Shaofeng Li, Minhui Xue, Benjamin Zi Hao Zhao, Haojin Zhu, Xinpeng Zhang, 2020. Invisible Attacks on Deep Neural Networks via Steganography and Regularization

[^9]: Kevin A. Zhang, Alfredo Cuesta-Infante, Lei Xu, Kalyan Veeramachaneni, 2019. SteganoGAN: High Capacity Image Steganography with GANs

[^10]: Isaac Corley, Jonathan Lwowski, Justin Hoffman, 2019. Destruction of Image Steganography using Generative Adversarial Networks

[^11]: Yang Yang, 2019. BASN — Learning Steganography with Binary Attention Mechanism

[^12]: Dongdong Hou, Weiming Zhang, Jiayang Liu, Siyan Zhou, Dongdong Chen, Nenghai Yu, 2018. Emerging Applications of Reversible Data Hiding

[^13]: Prafulla Dhariwal, Alex Nichol, 2021. Diffusion Models Beat GANs on Image Synthesis

[^14]: Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer, 2021. High-Resolution Image Synthesis with Latent Diffusion Models

[^15]: Weike You, Hong Zhang, Xianfeng Zhao, 2020. A Siamese CNN for Image Steganalysis

[^16]: Shashikala Channalli, Ajay Jadhav, 2009. Steganography An Art of Hiding Data

[^17]: Shumeet Baluja, 2017. Hiding Images in Plain Sight: Deep Steganography

[^18]: Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, Zhenyu Guan, ICCV 2021. HiNet: Deep Image Hiding by Invertible Network

[^19]: Xiyao Liu, Ziping Ma, Junxing Ma, Jian Zhang1, Gerald Schaefer, Hui Fang, CVPR 2022. Image Disentanglement Autoencoder for Steganography without Embedding

[^20]: Shao-Ping Lu, Rong Wang, Tao Zhong, Paul L. Rosin, CVPR 2021. Large-capacity Image Steganography Based on Invertible Neural Networks

[^21]: Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, Jian Zhang, CVPR 2022. Robust Invertible Image Steganography

[^22]: Jiren Zhu, Russell Kaplan, Justin Johnson, Li Fei-Fei, ICCV 2018. HiDDeN: Hiding Data With Deep Networks

[^23]: Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, In So Kweon, NeurIPS 2020. UDH: Universal Deep Hiding for Steganography, Watermarking, and Light Field Messaging

[^24]: Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, Kilian Q Weinberger, ICLR 2022. Fixed Neural Network Steganography: Train the images, not the network

[^25]: Mehdi Boroumand, Mo Chen, Jessica Fridrich, 2019. Deep Residual Network for Steganalysis of Digital Images