Texture attention-guided adaptive LSB steganography with U-Net model and optional GA tuning integrated with ablation study

Abstract

Image steganography involves hiding a secret image within a digital cover image, which remains invisible to ordinary viewers. The majority of current steganographic methods face an essential trade-off because higher Payload Capacity (PC) results in decreased visual undetectability and reduced resistance to detection and image distortions. Our proposed framework extracts the texture maps from the cover images, then generates the attention maps, and optionally uses Genetic Algorithm (GA) optimization to achieve maximum imperceptibility and robustness using content-adaptive and texture attention-guided methods. The system generates attention maps through the combination of deep texture maps extracted using ResNet50 blocks with distance transform maps, and then trains a U-Net model to predict these maps for new, unseen cover images. These attention maps guide adaptive bit allocation between 1 and 4 Bits Per Pixel (BPP). Next, we use the GA to optimize both attention threshold values and bit-allocation parameters, which enhances visual quality. The experimental results show that the Peak Signal-to-Noise Ratio (PSNR) values for all stego-images are between 55 dB and 61 dB. The corresponding Structural Similarity Index Measure (SSIM) values are close to 1.0. Moreover, our framework supports lossless recovery of secret messages embedded in stego-images. The Bit Error Rate (BER) between embedded secret messages and extracted secret messages is equal to 0. The corresponding SSIM values are equal to 1.0. To improve the embedding quality, we employ the GA to search for the optimal values of attention thresholds and adaptive bit allocation. The PSNR improvement of GA-based embedding for stego-images over attention-only embedding for stego-images is approximately 3.6–5.0 dB. Then, we conduct a series of ablation studies to evaluate the contributions of the attention predictor, i.e., a U-Net-based attention predictor, and the GA-based optimization for stego-images of high quality. The results demonstrate that both the attention predictor and the GA-based optimization are necessary to generate stego-images of high quality. We also conduct a computational analysis for embedding and extraction, and the results indicate that both embedding and extraction require only a few milliseconds per image after the one-time offline GA calibration. Finally, we conduct a security evaluation using three state-of-the-art Convolutional Neural Network (CNN)-based steganalyzers, i.e., Xu-Net, Ye-Net, and SR-Net, for stego-images generated by our framework. The detection performance of the three steganalyzers for stego-images generated by our framework is close to random guessing. The average Area Under the Curve (AUC) values for stego-images generated by our framework are approximately equal to 0.5.