Abstract
Recently, the generative adversarial network is the hotspot in research and industrial areas. Its application on data generation is the most common usage. In this paper, we propose the novel end-to-end framework to extend its application to data hiding area. The discriminative model simulates the detection process, which can help us understand the sensitivity of the cover image to semantic changes. The generative model is to generate the target image which is aligned with the original cover image. An attention model is introduced to generate the attention mask. This mask can help to generate a better target image without perturbation of the spotlight. The introduction of cycle discriminative model and inconsistent loss can help to enhance the quality of the generated target image in the iterative training process. The training dataset is mixed with intact images and attacked images. The mix training process can further improve robustness. Through the qualitative, quantitative experiments and analysis, this novel framework shows compelling performance and advantages over the current state-of-the-art methods in data hiding applications.
Introduction
Data hiding—the practice of embedding secret information (steganography) or signatures (watermarking) into digital media—requires a delicate balance between capacity, imperceptibility, and robustness. Traditional methods often rely on hand-crafted distortion functions to determine where to embed data. However, these methods and early deep learning attempts frequently struggle to maintain robustness or inadvertently place data in smooth areas where changes are easily detected by the human eye.
This research introduces Attention Based Data Hiding (ABDH), an end-to-end framework that leverages Generative Adversarial Networks (GANs) and an attention mechanism to automatically identify "inconspicuous" areas, such as textures and edges, for data embedding. By learning the semantic and structural complexity of a cover image, ABDH achieves superior imperceptibility while remaining resilient against common image processing attacks.
Methodology
The ABDH framework utilizes a sophisticated GAN architecture designed for high-fidelity data hiding and extraction:
- Attention Model ($G_A$): Utilizing a pre-trained ResNet-50 backbone, this module generates an "Attention Mask" that highlights complex structural areas (edges, textures) while avoiding smooth regions where the human eye is most sensitive to perturbations.
- Target Generator ($G_{TI}$): This component integrates the cover image, the secret image, and the attention mask to produce the "Target Image" (the stego-image).
- Secret Extractor ($G_{SI}$): A secondary generator responsible for the inverse process—reconstructing the hidden secret image from the target image.
- Dual Discriminators: A cover discriminator ($D_{CI}$) ensures the target image is indistinguishable from the original, while a cycle discriminator ($D_{SI}$) ensures the extracted secret is a faithful reconstruction.
- Inconsistent Loss ($L_{IC}$): This novel loss function ensures that the extraction process is unique to the target image. It prevents "false positive" extractions by requiring that the extractor produces a null or completely different output when applied to the original cover image.
The system is trained end-to-end using the COCO dataset with a "mixed training" strategy, exposing the network to simulated attacks like noise, JPEG compression, and cropping to build inherent robustness.
Results and Discussion
ABDH was evaluated against standard benchmarks (Set5 and Set14) and demonstrated significant improvements over traditional and deep-learning-based state-of-the-art (SOTA) methods.
Qualitative Analysis
Visual results demonstrate that the attention mechanism successfully concentrates the residual differences (the "noise" of the hidden data) into high-frequency areas like edges and textures. As shown in the comparison of world-renowned paintings, the target images remain virtually identical to the covers to the naked eye.
Quantitative Performance
- Imperceptibility: Target images achieved high Peak Signal-to-Noise Ratio (PSNR) values, typically exceeding 30-33 dB.
- Robustness: ABDH outperformed competitors like HiDDeN and SSGAN in recovery quality under attack. For instance, under "Salt and Pepper" noise, ABDH maintained a PSNR of 33.48, significantly higher than the 25.56 achieved by other deep learning methods.
- Capacity Trade-offs: The study quantified the relationship between the pixel-ratio (secret size vs. cover size) and performance, suggesting that for optimal security, the secret image should remain below half the size of the cover image.
Conclusion
The ABDH framework represents a significant advancement in data hiding by automating the discovery of optimal embedding regions through an attention mechanism. By simulating the detection process during training, the model learns to avoid features that would alert an adversary. Future research will focus on extending this mechanism to video data hiding, aiming to identify temporally consistent areas to prevent visual artifacts across frames.
Summary updated