Secure Multi-Secret Image Steganography Using Channel-Spatial Attention and Residual U-Net with Hybrid Perceptual Loss Optimization

Abstract

This work proposes a robust, deep learning-based multi-secret image steganography framework capable of securely embedding and recovering multiple secret images within a single cover image. The method introduces a Channel-Spatial Attention-enhanced U-Net architecture (Secret Decoder CBAM U-Net), designed to extract hidden information with high fidelity by leveraging convolutional block attention modules (CBAM) and residual connections for improved spatial feature preservation and gradient flow. A novel 6-bit embedding strategy is employed to encode two grayscale secret images simultaneously into a single cover image, enabling compact representation without significant quality degradation. The system is trained using a hybrid loss function that combines Mean Squared Error (MSE) and VGG-based Perceptual Loss, effectively balancing pixel accuracy and perceptual similarity. To assess robustness, the framework is subjected to adversarial and differential attacks, with evaluations based on PSNR, SSIM, UACI, NPCR, and SHA-256 hash integrity checks. While the model achieves good PSNR values above and high NPCR hash-based integrity validation reveals vulnerability to perturbations, indicating potential image alteration under attack conditions. This approach demonstrates promise in secure image communication and multimedia data hiding, combining multi-secret capacity, deep feature attention, and perceptual fidelity, while highlighting the need for future enhancement in resistance against adversarial interference.