[Paper Review] SSGAN: Secure Steganography Based on Generative Adversarial Networks

Haichao Shi, Jing Dong+3 more|arXiv (Cornell University)|Jul 6, 2017

Advanced Steganography and Watermarking TechniquesComputer Science14 references51 citations

TL;DR

The paper proposes a GAN-based framework for secure steganography with one generator and two discriminators to produce high-quality, robust covers for hiding information, achieving improved training stability and image quality.

ABSTRACT

In this paper, a novel strategy of Secure Steganograpy based on Generative Adversarial Networks is proposed to generate suitable and secure covers for steganography. The proposed architecture has one generative network, and two discriminative networks. The generative network mainly evaluates the visual quality of the generated images for steganography, and the discriminative networks are utilized to assess their suitableness for information hiding. Different from the existing work which adopts Deep Convolutional Generative Adversarial Networks, we utilize another form of generative adversarial networks. By using this new form of generative adversarial networks, significant improvements are made on the convergence speed, the training stability and the image quality. Furthermore, a sophisticated steganalysis network is reconstructed for the discriminative network, and the network can better evaluate the performance of the generated images. Numerous experiments are conducted on the publicly available datasets to demonstrate the effectiveness and robustness of the proposed method.

Motivation & Objective

• Motivate secure steganography by generating suitable image covers that resist detection.
• Develop a GAN-based framework that balances visual quality with information-hiding suitability.
• Improve convergence speed and training stability over traditional GAN approaches.
• Enhance the discriminative evaluation of generated covers via a strengthened steganalysis network.

Proposed method

• Use a novel GAN form with one generator and two discriminative networks.
• The generator focuses on visual quality of images suitable for steganography.
• Discriminators assess the suitability of generated images for information hiding and detect steganographic potential.
• Reconstruct and employ a sophisticated steganalysis network to better evaluate cover performance.
• Demonstrate improved convergence speed, training stability, and image quality on public datasets.

Experimental results

Research questions

• RQ1 Can the proposed SSGAN framework produce image covers that are both visually convincing and highly suitable for secure data embedding?
• RQ2 Does the alternative GAN formulation improve convergence speed and training stability compared to standard GAN setups?
• RQ3 Does the enhanced steganalysis-based discriminator better evaluate and guide the quality of covers for information hiding?

Key findings

• The approach yields improvements in convergence speed and training stability.
• Generated images show enhanced visual quality suitable for steganography.
• A strengthened steganalysis network offers better evaluation of cover performance.
• Experiments on public datasets demonstrate the effectiveness and robustness of the proposed method.