<p>Image steganography deals with the hidden transmission of information whereby a secret image is embedded within a cover image in a way that the secret image cannot be easily identified. In this research, we propose a steganographic system that combines edge-aware attention mechanisms using Holistically-Nested Edge Detection (HED) within deep learning frameworks to direct adaptive data embedding operations. The system starts by extracting edge maps using HED then converting them to an attention maps. These high-resolution distance-based attention maps direct adaptive bit embedding operations within cover images by adjusting the number of hidden bits per pixel through attention strength to maintain a balance between capacity and distortion. After that, we embed the secret image within the cover image based on these attention maps. In the embedding process, we use a custom adaptive Least Significant Bits (LSBs) strategy, which follows the predicted attention map that is generated from trained encoder-decoder CNN. On other hand, we optimize the embedding process using a genetic algorithm (GA) to enhance the embedding process through adjusting the threshold values of attention map rules. In this work, we use two datasets (USC-SIPI as secret images and Boss Base as cover images) to test and train our stenographic system. We assessed the performance of our optimized deep learning model based on various performance metrics like Mean Square Error (MSE), Peak Signal to Noise Ratio (PSNR), Mean Absolute Error (MAE), Structural Similarity Index (SSIM), Image Fidelity (IF), Payload Capacity (PC), Bit per pixel (BPP), Xu-Net, Ye-Net, and RS steganalysis analysis. The experimental results show that the PSNR value of 60.72–61.20 dB emerges from 0.1 BPP. The SSIM values of 0.9995–0.9996 emerge from this level. The PSNR values reach 58.27 dB and 55.16 dB when the BPP ratio reaches 0.195 and 0.397 respectively while maintaining a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textrm{BER}=0/\textrm{BR}=100\%\)</EquationSource> </InlineEquation> throughout. The steganalyzers fail to detect the hidden data because XuNet and YeNet achieve AUC values between 0.47 and 0.53 while RS analysis produces incorrect embedding rate predictions between 0.63 and 0.65. The system maintains its robustness against salt-and-pepper noise because it achieves BER values of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\((\textrm{BER}\approx 0.005\ \text {at } p=0.01;\ \approx 0.051\ \text {at } p=0.10)\)</EquationSource> </InlineEquation> whereas cropping, additive Gaussian noise, and JPEG compression drive extraction toward random <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\((\textrm{BER}\approx 0.5)\)</EquationSource> </InlineEquation>. Runtime is modest (per image): <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\text {embedding}\ \approx 0.04\text {-}0.17\,\textrm{s},\ \text {extraction}\ \approx 0.02\text {-}0.11\,\textrm{s}\)</EquationSource> </InlineEquation>. The proposed method achieves a good trade-off between capacity and imperceptibility and security at low computational cost while keeping the remaining vulnerabilities under cropping and compression attacks.</p>

Secure edge-guided adaptive image steganography using HED-based attention maps and CNN

期刊论文
出版模式：

开放获取
发表日期：

2025年12月2日

查看全文

Rana Alrawashdeh,
Sultan Almuhammadi,
Mahmood Niazi,
Mufti Mahmud,
Md Mahfuzur Rahman

摘要

Image steganography deals with the hidden transmission of information whereby a secret image is embedded within a cover image in a way that the secret image cannot be easily identified. In this research, we propose a steganographic system that combines edge-aware attention mechanisms using Holistically-Nested Edge Detection (HED) within deep learning frameworks to direct adaptive data embedding operations. The system starts by extracting edge maps using HED then converting them to an attention maps. These high-resolution distance-based attention maps direct adaptive bit embedding operations within cover images by adjusting the number of hidden bits per pixel through attention strength to maintain a balance between capacity and distortion. After that, we embed the secret image within the cover image based on these attention maps. In the embedding process, we use a custom adaptive Least Significant Bits (LSBs) strategy, which follows the predicted attention map that is generated from trained encoder-decoder CNN. On other hand, we optimize the embedding process using a genetic algorithm (GA) to enhance the embedding process through adjusting the threshold values of attention map rules. In this work, we use two datasets (USC-SIPI as secret images and Boss Base as cover images) to test and train our stenographic system. We assessed the performance of our optimized deep learning model based on various performance metrics like Mean Square Error (MSE), Peak Signal to Noise Ratio (PSNR), Mean Absolute Error (MAE), Structural Similarity Index (SSIM), Image Fidelity (IF), Payload Capacity (PC), Bit per pixel (BPP), Xu-Net, Ye-Net, and RS steganalysis analysis. The experimental results show that the PSNR value of 60.72–61.20 dB emerges from 0.1 BPP. The SSIM values of 0.9995–0.9996 emerge from this level. The PSNR values reach 58.27 dB and 55.16 dB when the BPP ratio reaches 0.195 and 0.397 respectively while maintaining a \(\textrm{BER}=0/\textrm{BR}=100\%\) throughout. The steganalyzers fail to detect the hidden data because XuNet and YeNet achieve AUC values between 0.47 and 0.53 while RS analysis produces incorrect embedding rate predictions between 0.63 and 0.65. The system maintains its robustness against salt-and-pepper noise because it achieves BER values of \((\textrm{BER}\approx 0.005\ \text {at } p=0.01;\ \approx 0.051\ \text {at } p=0.10)\) whereas cropping, additive Gaussian noise, and JPEG compression drive extraction toward random \((\textrm{BER}\approx 0.5)\) . Runtime is modest (per image): \(\text {embedding}\ \approx 0.04\text {-}0.17\,\textrm{s},\ \text {extraction}\ \approx 0.02\text {-}0.11\,\textrm{s}\) . The proposed method achieves a good trade-off between capacity and imperceptibility and security at low computational cost while keeping the remaining vulnerabilities under cropping and compression attacks.