HGAN-U-Net: Hybrid Generative Adversarial Network-U-Net Approach for Multi-Domain Text Steganography

2.1 Convolutional neural network-based deep steganography

An important shift occurred with the introduction of CNN-based image steganography research, which replaced manually crafted spatial-domain rules such as LSB replacement with adaptive frequency-domain rules such as HUGO. Without the need for a manually created embedding rule, Baluja [9] demonstrated how two convolutional encoder-decoder networks could immediately learn an embedding rule to embed and extract a full-size color image from the training dataset. To achieve a significantly greater embedding payload than one bit per pixel, Zhang et al. [10] further expanded on this concept by inventing SteganoGAN, which included adversarial training and dense convolutional blocks. Although the embedding capacity of these early models was significantly increased, most of the resulting stego images had lower peak signal-to-noise ratio (PSNR) values and were not particularly resistant to statistical steganalysis methods. Additionally, only one dataset (ImageNet) was used to train both architectures, and the assessment was conducted almost exclusively within this domain, with less focus on cross-domain generalization [1].

2.2 Adversarial GAN-based steganography

This provided the foundation for subsequent work that specifically used adversarial training to increase learned steganalyzers' adversarial resilience. To avoid this, Fu et al. [11] developed a residual encoder–decoder Generator that was paired with a discriminator to distinguish between stego and natural cover images, resulting in less distortion than earlier GAN-based techniques. Ramandi et al. [12] introduced a novel adaptive architecture of VidaGAN that includes dedicated loss terms and a distinct critic network to achieve an ideal trade-off between embedding capacity and visual transparency. These adversarially trained techniques greatly reduced detectability, but they neglected resilience against channel distortion in favor of security [11, 12]. This implies that recovering the encoded payload is typically not achievable if a stego image is compressed using JPEG after transmission. Additionally, a number of these systems were designed to hide a complete secret picture rather than a bit stream of text, which is less crucial than sending text, as it can handle certain pixel-level faults in an image.

2.3 Robustness and error-correction approaches

A further line of inquiry was thus focused on resilience to actual channel noise and lossy compression. To simulate JPEG-like distortion, cropping, and blurring, Zhu et al. [13] placed a differentiable noise layer between the encoder and decoder during training. Later, Jing et al. [14] introduced an invertible neural network termed HiNet for concealing and extracting using an invertible transform, which yields superior fidelity for a tiny perturbation. By employing adaptive error-correcting codes and embedding-domain selection to further minimise residual bit errors, Duan et al. [15] safeguarded the payload against lossy JPEG recompression. Despite these advancements, most robustness-focused models sacrifice some peak signal-to-noise for resilience when compressed [13-15]. Moreover, many of these methods fail to achieve near-perfect text recovery, and only a few have been evaluated over a diverse array of heterogeneous cover-image domains to verify genuine generalization [1].

2.4 Cross-domain generalization methods

Rather than focusing on a particular benchmark, a fourth, relatively understudied issue was generalization across various cover-image domains. Stegano-CNN, an early convolutional architecture specifically designed to maintain embedding quality across a range of image textures, was developed by Duan et al. [16]. Steg-GMAN was presented by Huang et al. [17], who trained the Generator against numerous steganalyzers at once and showed consistent security when tested on multiple separate datasets. Instead of using a single source, Zhou et al. [18] used a purposefully mixed pool of benchmark datasets and real-world photos to train a latent-space embedding system. Compared to single-dataset baselines, our mixed-domain architecture significantly increased resilience to unseen cover statistics. However, comprehensive cross-domain evaluation is relatively understudied because the majority of current frameworks continue to provide results on a single dominating benchmark [19].

Propose a hybrid adversarial framework to overcome the intrinsic drawbacks of current GAN-based steganography techniques, which usually prefer embedding capacity at the price of visual quality and suffer from poor cross-domain generalization. This method is trained on a composite multi-domain dataset (BOSSBase, DIV2K, COCO, and CelebA) to guarantee strong generalization, in contrast to other methods that train on single-domain datasets. Additionally, the proposed method deliberately modifies the strategic paradigm by sacrificing the ability to achieve extreme visual imperceptibility to improve statistical security against steganalysis and to guarantee deterministic message recovery via an integrated ECC in the interest of optimising the raw payload (Bpp).

The experimental findings for the proposed system are presented in this section from several complementary perspectives. Each of the four connected subsections that make up the conversation offers a pertinent assessment of the whole topic. The experimental settings are explained in the first subsection. The second subsection then goes into the training dynamics and convergence analysis. The final subsection, visual quality and message recovery, provides a comparison with recent relevant work. The ablation study is the final section.

4.1 Experimental settings

The proposed framework was developed on a Linux-based system using the PyTorch deep learning library. All experiments were run on a single NVIDIA Tesla T4 GPU from the Kaggle platform. The network has a lightweight architecture with about 812,000 trainable parameters to achieve computational efficiency. Adam was used as an optimizer, starting with an initial learning rate of 0.001 for both networks. The strict weight decay was set to 1e-5 to avoid overfitting in the adversarial training. Moreover, a gradient clipping technique was used to stabilise the GAN training process, with a maximum norm of 1.0. The total training time for all 30 epochs was ~3.75 hrs. The random seed was set to 42 to ensure full repeatability of the reported results. Furthermore, deterministic algorithms for cuDNN were activated in all training and evaluation steps.

To produce a balanced multi-domain composite dataset and reduce the notable class imbalance between the larger MS-COCO (163k images) and CelebA (405k images) datasets in comparison to the smaller BOSSBase and DIV2K datasets, a rigorous sample restriction was used. Specifically, 2000 images from BOSSBase, 1800 from DIV2K (its whole test set), 2000 from MS-COCO, and 2000 from CelebA were sampled to generate a balanced total of 7,800 images. The complete dataset was randomly shuffled and split into 6240 training images (80%), 1170 validation images (15%), and 390 testing images (5%) to ensure no overlap between the training and testing sets. During preprocessing, each image was resized to 256 × 256 × 3 using bicubic interpolation. Only the training split was subjected to data augmentation, which included random horizontal flips, vertical flips, and 90° rotations, in order to improve spatial invariance and avoid overfitting. This model was trained for 30 epochs using an Adam optimizer and an initial learning rate of 1 × 10⁻³. In the lengthy adversarial training, a Cosine Annealing scheduler was used to obtain stable convergence. The weights of the loss functions were empirically optimized, and a batch size of 16 was employed.

Throughout training, a robustness layer was included to model Gaussian noise and JPEG compression. The secret text was also encoded using a 3× repetition ECC before being embedded.

The proposed system is analyzed according to four primary features, each of which is evaluated using particular metrics to fully evaluate the method: (1) Visual quality—PSNR and SSIM; (2) Embedding capacity—Effective BPP; (3) Recovery accuracy—AccTxt; (4) Detection accuracy (security)—AUC and Chi-square p-value.

4.2 Training dynamics and convergence analysis

Training loss, adversarial loss, PSNR, and bit accuracy were displayed across 30 epochs in Figure 2. The validation loss decreased significantly throughout the first seven epochs, reaching its lowest value of 0.0499 at epoch 7. The accuracy of the maximum bit. Thus, epoch 7 was the model's most balanced operating point. In most epochs, the discriminator loss remained close to zero, indicating that steganalyzer had difficulty distinguishing between cover and stego images. There was a sudden change following. The bit accuracy had reduced to 22.49%, with a validation loss of 4.5448. The network gradually developed, with the number of accurate bits in the output reaching 92.15% by epoch 30. This pattern showed that the model would generalize to more robust channel perturbations, even if it was briefly "disturbed”.

2.png

Figure 2. Training curves demonstrating bit accuracy, peak signal-to-noise ratio (PSNR), and loss stability across 30 epochs

4.3 Visual quality and message recovery

The complete assessment of 100 images selected from each of the four datasets is displayed in Table 4. The method's average PSNR and SSIM values are 49.74 dB and 0.9987, respectively, both of which are significantly higher than the range deemed visually imperceptible. The average absolute difference per pixel remained below 328.45, and the mean squared error remained below 0.000012. The extraction accuracy was 99.68%, with a bit error rate of 0.32%. The most stringent requirement would be that the text be retrieved precisely, just as with text messages. Of the 100 messages tested, 86.00% were successful. This disparity resulted from the fact that the reconstructed text might alter if a single bit in the reconstructed bit stream was wrong. The triple copying used by the error resilience code produced the same effective embedding rate (0.0052 bits/pixel).

Table 4. Over 100 held-out images from the four-dataset validation pool were averaged for the comprehensive evaluation report

Table 5. Compared to other current information-concealing methods

Table 5 shows that, in terms of optimal performance in stego-image quality and embedded-data security, the proposed solution outperforms state-of-the-art deep learning-based data concealment techniques. The findings imply that contemporary models, such as VidaGAN [12] and Ji et al. [20], aim to maximize the embedding space. At the expense of visual quality, these techniques can achieve embedding capacities between 1.00 and 3.90 bits per pixel, with SSIM values of 0.95 to 0.98 and PSNR values of no more than 41.50. In contrast, the proposed approach has an entirely different design philosophy. With a maximum PSNR of 49.74 dB and an SSIM of 0.9987, the model presented in this research produced outstanding visual results, indicating that neither the human eye nor analytic software can detect changes in the original image. This decrease in embedding capacity to 0.0052 bits per pixel is a deliberate design decision to obtain better visual quality. This reduction is directly attributable to the addition of ECC to ensure total strength in text recovery rather than model restrictions. The integrity of retrieval is significantly more crucial for text-sensitive applications than payload size, and visual concealment is also crucial. Therefore, this research demonstrates that while the suggested model is not competitive in data throughput, it performs admirably in concealment quality and robustness of retrieval.

It is crucial to remember that this sort of numerical comparison is challenging. Initially, single-domain datasets are used to evaluate the baseline models. In contrast to previous frameworks, ours is trained and assessed using a multi-domain dataset. These other architectures could not be operated with our setup and computational needs. Consequently, we employed the measures they reported as most effective. We selected baseline values from datasets such as DIV2K and MS-COCO that coincide with our composite set. This guarantees a benchmark that is reasonable from a scientific standpoint. The comparison also focuses on the strategic equilibrium between image quality and capacity. Different techniques emphasize payload throughput, but here we trade capacity for excellent imperceptibility and cross-domain generalization.

The other methods, such as Liu et al. [22] and Sanjalawe et al. [23], achieve higher PSNR values and embedding capacities, but this is a deliberate architectural trade-off and does not stem from the limitations of the model. First, they are trained and tested on single-domain datasets, whereas our framework is specifically designed to deal with multi-domain generalization (BOSSBase, DIV2K, COCO, CelebA), which makes embedding harder. Secondly, they lack an ECC or an adversarial robustness layer. As mentioned, we knowingly compromise a bit of raw capacity (from 0.0188 Bpp to 0.0052 Bpp via 3 times bit repetition) to ensure deterministic 100% recovery of text under noisy channel conditions. Moreover, our model does not rely solely on visual fidelity, as most existing models do, but employs an active adversarial training mechanism against a Steganalyzer to achieve better statistical security (AUC = 0.5573). Therefore, the proposed framework doesn't compete on raw data throughput; instead, it offers a sweet spot for cross-domain generalization, extreme robustness, and anti-detectability.

Scientific circles widely acknowledge that a very low embedding capacity is directly correlated with extremely high visual quality. In scientific circles, it is widely acknowledged that a very low embedding capacity is directly related to an extremely high visual quality. The network needs to change fewer pixels because the payload uses fewer bits per pixel. As a result, the minimal modification rate yields extremely high PSNR and SSIM values. So, our approach does not aim to assert superiority in visual quality over high-capacity steganography models. Rather, we deliberately focus on extreme visual imperceptibility and strong data recovery rather than raw data throughput. For such specific, highly sensitive practical situations, this architectural idea is well suited to the proposed method. For instance, complete fidelity is required when transmitting brief text messages containing vital information, such as encryption keys or secret locations. For these sorts of uses, any bit error or visible artefact will taint the entire covert channel. We, however, admit that this technique is not, compared to high-capacity steganography methods, capable of providing optimal performance. Approaches based on raw capacity, such as ones based on invertible neural networks, can embed full-resolution images. Our framework is not tailored to support bulk data transfer or high-bandwidth multimedia hiding applications. If the capacity exceeds the current limit, its structural integrity will certainly be compromised.

4.4 Cross-domain generalization analysis

To thoroughly demonstrate the validity of the proposed framework’s cross-domain generalization, the trained model was evaluated on each dataset separately. The performance differences among the four image types are presented in Table 6.

Table 6. Performance variations across different image domains

As shown in Table 6, the experimental results clearly demonstrate high stability and consistency across all four datasets tested. The texture distribution is the reason for slightly higher PSNR values for BOSSBase and CelebA. The BOSSBase dataset, for example, has rich, high-frequency textures that are useful for hiding subtle embedding modifications. In contrast, CelebA has smooth skin parts, and the shape edges are sharp, which makes its embedding highly structured. The graphical fidelity and text recovery rates of the two are generally good, albeit with slight differences between them in DIV2K and MS-COCO. Most significantly, our detailed analysis shows that none of the image domains has any significant performance loss. The text recovery accuracy (AccTxt) shows a highly consistent profile across all assessed domains. The high attention stability is proof of the successful adaptation of the integrated AG to very complex images. This means that the proposed model can be embedded in any visual domain without overfitting.

4.5 Impact analysis of the error correction code module

To comprehensively assess the influence of the ECC, a more detailed analysis of the parameters of its operation was carried out. First, the bit error rate (BER) was analyzed in detail before and after implementing the proposed ECC module. The raw bit error rate for the neural network remains at 0.32% all the time when it lacks the ECC module. But when this 3× repetition code is integrated into the system, this effective error rate drops to about 0.003%. This is a substantial decrease to make the chance of double-bit errors ruining a character as small as possible. Second, we thoroughly examined various repetition strategies to ensure that we had the best operational/computational compromise. A 2-rep approach does not have the mathematical foundation needed to correct bit errors with the majority vote technique. Furthermore, a 5× repetition setting provides greater robustness, but at the expense of an unnecessary 80% capacity loss. Hence, the strategy of three times was chosen as the best trade-off between error-correcting capacity and payload retention. Finally, this analysis explicitly calculates the capacity loss due to the ECC module. The embedding capacity of the network when no error correction is used is 0.0156 bits per pixel. The advantage of the 3× repetition code is that it provides an effective capacity of 0.0052 bits per pixel. This is an intentional 66.6% capacity compromise to ensure deterministic text recovery in the presence of noise.

4.6 Security and anti-detectability analysis

The steganalyzer test dataset was meticulously constructed to assess anti-detection performance. It was made up of an equal number of cover and stego images in a flawlessly balanced blend. The Receiver Operating Characteristic curve was used to calculate the Area Under the Curve (AUC). To assess detection capabilities, this curve plots the true positive rate versus the false positive rate. The steganalyzer is only useful when the AUC value is close to 0.5, indicating that the stego images are statistically undetected. Additionally, the statistical distribution of Least Significant Bit pairs was particularly assessed using the Chi-square test. Anomalies in the Pair of Values frequencies brought on by secret data embedding are found using this traditional technique. In terms of detection methods, we looked at both a traditional statistical detector and a deep learning-based steganalyzer. This thorough dual-model approach ensures strong validation against both contemporary neural and conventional statistical steganalysis approaches. The suggested model demonstrated outstanding security under this assessment approach. With an AUC of 0.5573, the internal deep steganalyzer came close to the optimal random guessing threshold. Additionally, the Chi-Square statistical test yielded a P-value of 0.286, which is well above the 0.1 detection threshold. Lastly, the t-distributed stochastic neighbor embedding (t-SNE) visualization shows a significant degree of overlap between the deep features of cover and stego images, as shown in Figure 3, offering indisputable visual evidence of the model's anti-detectability.

3.png

Figure 3. t-SNE visualization of cover vs stego features

4.7 Ablation study

To confirm the requirement of each suggested component, we conducted an inference-time ablation study in which we systematically turned off elements of the system and observed the performance loss. A notable decline in visual quality (PSNR dropped from 49.74 dB to 41.32 dB) was observed when AG were removed from the skip connections, as shown in Table 7. This demonstrates the significance of AG in guiding the embedding process toward intricate texture regions and avoiding noticeable artifacts. Furthermore, the bit extraction accuracy dropped to 81.45% when the Receiver network's multi-scale extraction heads were disabled, suggesting that spatial redundancy across many scales is necessary for message recovery. Ultimately, it was found that the neural network provides high bit accuracy and that the ECC is a crucial component that guarantees text recovery at 100% accuracy; otherwise, it falls to 6.50%. These findings demonstrate the structural need of each connected element.

Table 7. Ablation study results on the multi-domain composite dataset

Contribution of the SPP module: we replaced it with a conventional convolutional block. The results indicate that the SPP module plays a crucial role in capturing global contextual information during the embedding process. Additionally, we evaluated the impact of the robustness layer by removing it during training. As expected, the absence of simulated distortions slightly increased PSNR for clean images to 51.40 dB while maintaining 100% text recovery. However, this configuration lacked robustness under practical conditions. When subjected to Gaussian noise or JPEG compression, the bit recovery accuracy decreased significantly without the robustness layer. These results demonstrate that the robustness layer is essential for ensuring message recovery under common image distortions.