Novel Generative Adversarial Network with Enhanced Attention Mechanism for Robust Image Steganography

The evolution of steganography has been fueled by advancementsin machine learning, yet achieving an optimal balance between imperceptibility, robustness, and embedding capacity remains a challenge.This paper introduces an attention-enhanced generative adversarialnetwork (AGAN) framework for image steganography. By incorporating spatial attention modules into the generator and discriminatornetworks, the proposed method identifies and utilizes regions of thecover image that are less perceptible to human vision for embeddingsecret data. This strategy minimizes perceptual distortion while maximizing robustness against steganalysis and image processing operations. Additionally, a hybrid loss function is introduced, combiningadversarial, reconstruction, and attention-guided perceptual losses toensure high-quality stego-images and reliable data recovery. Experimental evaluations on benchmark datasets demonstrate that AGANoutperforms state-of-the-art techniques in Peak Signal-to-Noise Ratio(PSNR), Structural Similarity Index Measure (SSIM), and resistanceto advanced detection tools. This work not only advances the capabilities of GAN-based steganography but also underscores the potentialof attention mechanisms in secure digital communication.