Recent Advances of Image Steganography with Generative Adversarial Networks

Jia Liu, Yan Ke, Lei Yu, Zhuo Zhang, Jun Li, Peng Luo, Minqing Zhang, Xiaoyuan Yang

Abstract

In the past few years, the Generative Adversarial Network (GAN) which proposed in 2014 has achieved great success. GAN has achieved many research results in the field of computer vision and natural language processing. Image steganography is dedicated to hiding secret messages in digital images, and has achieved the purpose of covert communication. Recently, research on image steganography has demonstrated great potential for using GAN and neural networks. In this paper we review different strategies for steganography such as cover modification, cover selection and cover synthesis by GANs, and discuss the characteristics of these methods as well as evaluation metrics and provide some possible future research directions in image steganography.

Introduction

Generative Adversarial Networks (GANs) are revolutionizing image steganography—the art of hiding secret data within digital images. By moving from manually designed rules to data-driven adversarial learning, GANs are enabling more secure, robust, and creative ways to achieve covert communication. This paper surveys the transition from manual, rule-based embedding to automated, data-driven approaches, categorizing these advances into three main strategies: cover modification, selection, and synthesis.

The Adversarial Framework

Traditional steganography faces a constant "arms race" against steganalysis (the detection of hidden messages). Historically, researchers manually designed "distortion functions" to identify which pixels in an image could be changed with the least risk of detection. However, as steganalysts began using sophisticated machine learning to spot these manual patterns, traditional methods became vulnerable.

The general framework of image steganography involving Alice (sender), Bob (receiver), and Wendy (steganalyst).

Steganography is inherently a game between Alice (sender) and Wendy (warden). GANs are the first tools to model this relationship directly in the training objective function. This shift from "hand-crafted" features to "learned" features allows steganography to adapt to the specific statistical distribution of the data it is hiding in.

Methodologies in GAN-based Steganography

The authors categorize GAN-based steganography into three fundamental architectures:

1. Cover Modification

GANs are used to improve existing images. This includes using a generator to create "secure" cover images (e.g., SGAN, SSGAN) or using GANs to automatically learn the best "distortion probability matrix" (e.g., ASDL-GAN), effectively teaching the system where it is safest to hide data.

2. Cover Selection

Instead of modifying an image, the system uses a GAN to establish a mapping between a secret message and an existing, unmodified natural image (e.g., the GSK method). This is perfectly secure against statistical analysis because the image itself is never altered.

3. Cover Synthesis

This is the most innovative use of GANs. The stego-image is generated directly from a "black box" or a generative model, rather than modifying an existing file.

The process of cover synthesis, distinguishing between manual design for unnatural images and data-driven design for natural images.
  • Unsupervised (SWE): Mapping messages directly into the random noise vector that feeds the generator.
  • Semi-supervised (ACGAN): Hiding messages within image labels or categories.
  • Constraint Sampling (GSS): Iteratively refining a generated image until it satisfies both "realness" and the hidden message constraints.
  • CycleGAN: Using image-to-image translation to create reversible steganography.

Key Insights and Results

  • The "No Original" Advantage: In cover synthesis, there is no "original" cover image for a warden to compare against. This renders many traditional "differential" steganalysis attacks completely useless.
  • Security Performance: Methods like SSGAN achieved error rates for steganalysts as high as 0.72, meaning the detector was performing worse than random guessing in some scenarios.
  • Robustness: Synthesis methods have demonstrated 100% extraction accuracy even after the stego-images were subjected to JPEG compression and contrast enhancement.
  • Capacity vs. Security: While modification methods have high payloads, synthesis methods currently have much lower relative capacities (e.g., 1.49e-3 to 1.10e-2 bytes/pixel), but offer significantly higher security.

Future Directions

Despite these advances, several challenges remain. GAN-based synthesis still lags behind traditional modification methods in terms of capacity. Furthermore, while a GAN image might hide a message well, it may still be detectable as a "generated" image by image forensics tools. Future work must bridge the gap between steganalysis (finding the message) and image forensics (finding the GAN). Additionally, the field needs better standardized metrics for "realness" and "security" in a generative context, as traditional metrics like PSNR are less applicable to synthesized content.

Summary updated

Abstract

In the past few years, the Generative Adversarial Network (GAN) which proposed in 2014 has achieved great success. GAN has achieved many research results in the field of computer vision and natural language processing. Image steganography is dedicated to hiding secret messages in digital images, and has achieved the purpose of covert communication. Recently, research on image steganography has demonstrated great potential for using GAN and neural networks. In this paper we review different strategies for steganography such as cover modification, cover selection and cover synthesis by GANs, and discuss the characteristics of these methods as well as evaluation metrics and provide some possible future research directions in image steganography.
BibTeX
@inproceedings{liu2020recent,
  author = {Jia Liu and Yan Ke and Lei Yu and Zhuo Zhang and Jun Li and Peng Luo and Minqing Zhang and Xiaoyuan Yang},
  title = {{Recent Advances of Image Steganography with Generative Adversarial Networks}},
  booktitle = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2983175},
}
Figure