Enhance Steganography Payload based on Deep Learning Model

Abstract

Image steganography is the process of hiding confidential information within digital images by introducing subtle modifications that are generally imperceptible to human vision. However, conventional steganographic techniques often encounter challenges related to limited data-hiding capacity, reduced robustness, and vulnerability to various forms of analysis or attack This paper proposes image steganography based on generative adversarial networks, called IM-GAN, used to improve the capacity of the image steganography framework. The proposed work includes three steps: first, the encoder steps used for embedding secret data into cover images. The second is a decoder for secret message recovery. Thirdly, a critic for visual quality assessment and a noise layer to enhance robustness against attacks. Through the combination of densely connected convolutional features, a Residual Kolmogorov Arnold Network (RKAN)-based convolutional layer, and a multi-objective loss function, IM-GAN enhances hiding capacity while preserving high visual quality and maintaining strong security. Three different datasets are used (DIV2K, MS COCO), along with BossBase 1.01 (512x512 pixels), to evaluate the results. The IM-GAN attains an embedding capacity of up to 4 bits per pixel (bpp) with high decoding accuracy and minimal perceptual distortion, even under Gaussian noise interference. Different parameters used to evaluate the proposed work demonstrate enhancement in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Human Visual System (HVS), SRNet, XuNet, and Ye-Net, while still achieving competitive performance even when trained using a relatively small dataset.