Three-phased multi-scale residual-dense modified-U-Net architecture for deep image steganography

Abstract

Recent advancements in deep image steganography have shown promise, yet many existing approaches fail to address the inherently lossy nature of neural networks, limiting their effectiveness. To overcome this limitation, we propose a novel architecture that integrates residual dense multi-scale JBs and FEBs within a three-phased U-Net framework, composed of three interconnected sub-networks: the secret encoder, encoder, and decoder. This modular design enhances data encoding and facilitates easy adaptation to different datasets, offering greater flexibility. By incorporating multi-scale processing to minimize information loss and leveraging high inter-connectivity to improve feature hiding. Additionally, a custom loss function, combining seven distinct components, is employed to guide the model’s learning process. The effectiveness of this architecture is demonstrated on CIFAR-10, CIFAR-100, and CelebA datasets, where it achieves an average improvement of 4.76 in peak signal-to-noise ratio (PSNR) and 0.0292 in structural similarity index measure (SSIM), which highlight the potential proposed architecture to elevate the field of image steganography.