GAN-Based Image Steganography and Steganalysis Framework Using Adversarial Learning

Secure communication in modern digital networks requires intelligent information-hiding mechanisms capable of resisting deep-learning-based attacks. Traditional steganography methods such as LSB, DCT, or wavelet-based embedding fail against modern statistical and CNN-based steganalysis. This research presents a robust, adversarial trained GAN-based steganography system that embeds secret text inside images while resisting detection by a dedicated attacker network. The proposed architecture consists of a Generator– Extractor pair, a Discriminator, and a Steganalysis Attacker, trained in a multi-phase curriculum for improved imperceptibility, message recovery, and attack resilience. The system is trained on the STL10 dataset using PyTorch and evaluated using PSNR, SSIM, and BER metrics. Results show high visual quality, low distortion, and reliable message extraction even after JPEG compression and noise injection. This work demonstrates a secure, adaptive, and high-capacity steganography framework suitable for modern digital forensics, covert communication, and privacy-preserving applications.