Abstract
In the past few years, the Generative Adversarial Network (GAN), which proposed in 2014, has achieved great success. There have been increasing research achievements based on GAN in the field of computer vision and natural language processing. Image steganography is an information security technique aiming at hiding secret messages in common digital images for covert communication. Recently, research on image steganography has demonstrated great potential by introducing GAN and other neural network techniques. In this paper, we review the art of steganography with GANs according to the different strategies in data hiding, which are cover modification, cover selection, and cover synthesis. We discuss the characteristics of the three strategies of GAN-based steganography and analyze their evaluation metrics. Finally, some existing problems of image steganography with GAN are summarized and discussed. Potential future research topics are also forecasted.
Introduction: The Steganographic Challenge
Steganography is the practice of concealing a secret message within an ordinary medium, such as a digital image, to facilitate covert communication. Unlike cryptography, which hides the meaning of a message, steganography hides the very existence of the message. The primary adversary is the "warden" or steganalyst, whose goal is to detect statistical or visual anomalies indicating a hidden payload.
Traditional steganography relies on the "Steganographic Triangle," balancing security (undetectability), capacity (payload size), and robustness (resistance to processing). Conventional methods typically modify cover image pixels based on hand-crafted rules. However, these modifications often leave detectable statistical footprints. Generative Adversarial Networks (GANs) offer a paradigm shift by moving from manual rule-setting to learning data distributions.
GAN-Based Steganography Strategies
The research identifies three primary strategies for integrating GANs into steganography:
1. Cover Modification (GAN-CMO)
In this approach, GANs replace human-designed embedding rules. Instead of minimizing simple pixel distortion, models like ASDL-GAN use a generator to produce a "probability map" identifying the safest regions for data hiding. A discriminator acts as a proxy for a steganalyst, providing feedback that helps the generator refine its embedding strategy.
2. Cover Selection (GAN-CSE)
Cover selection is a "no-modification" strategy. GANs are utilized to establish a complex mapping between secret messages and a pre-existing database of natural images. To communicate, the sender transmits a completely unmodified natural image that corresponds to the intended message. Because the image remains untouched, traditional steganalysis techniques are rendered ineffective.
3. Cover Synthesis (GAN-CSY)
This strategy involves generating the stego-image from scratch.
- Supervised Synthesis: Uses encoder-decoder structures (e.g., SteganoGAN) to embed data during the image generation process.
- Unsupervised Synthesis: Maps messages directly to the latent noise vectors used by GANs. The resulting image is the message.
- Semi-supervised Synthesis: Employs class labels (e.g., in ACGAN) to represent specific message components.
Datasets and Evaluation
Research in this field leverages several benchmark datasets:
- BOSS (Break Our Steganographic System): The standard for testing detectability in traditional and modern steganography.
- CelebA: Used extensively for facial image synthesis and modification.
- MNIST & LFW: Common for proof-of-concept synthesis and selection tasks.
- Div2k & COCO: High-resolution datasets for testing steganography in complex, real-world scenarios.
Key Insights and Results
The transition from distortion minimization to distribution matching is the most significant contribution of GANs. By minimizing the distance between the probability distributions of cover and stego-images, GANs theoretically achieve higher levels of undetectability.
- Security: GAN-CMO methods like ASDL-GAN have successfully discovered embedding strategies that outperform established human-designed algorithms like HILL or S-UNIWARD.
- Capacity: Synthesis methods like SteganoGAN provide high payloads while maintaining visual quality, though they remain vulnerable to specialized machine-learning steganalysis.
- Robustness: Innovations like StegaStamp demonstrate that GANs can hide data robustly enough to survive physical transmission, such as printing and re-photographing.
Limitations and Future Directions
Despite significant advances, several challenges remain:
- Capacity-Security Trade-off: High-capacity synthesis often introduces visual artifacts that compromise security.
- Metrics: Traditional metrics like PSNR are insufficient for measuring steganographic security. The field is moving toward metrics like FID (Fréchet Inception Distance) to better evaluate the "naturalness" of stego-images.
- Computational Efficiency: The high resource requirements for training GANs limit their deployment on edge devices.
- JPEG Domain: Most current research focuses on the spatial domain, whereas most internet traffic uses JPEG compression. Future research must address the DCT (Discrete Cosine Transform) domain to be practical for web-based communication.
Summary updated