GAN-based image protection with reversible data concealing via secret key extraction

The swift growth of artificial intelligence in digital media has raised growing concerns regarding image misuse, unauthorised recognition, and breaches of ownership rights. Current methodologies seldom attain traceability, imperceptibility, robustness, and full reversibility within one framework. This research presents a unified image protection system that combines transform domain steganography, secret key guided reversible data hiding (RDH), and adversarial perturbations produced by Generative Adversarial Networks (GANs). The GAN component generates subtle perturbations that successfully deceive deep neural network classifiers, thus maintaining privacy, while imperceptible steganographic embedding guarantees the secure concealment of copyright information. The embedded payload functions as a fragile authentication fingerprint: any post-embedding image modification disrupts payload recovery, providing evidence of tampering. The RDH mechanism facilitates precise restoration of both the original image and the embedded payload, thereby ensuring complete reversibility. Comprehensive experiments on benchmark datasets indicate that the proposed framework reliably maintains high visual quality, with PSNR values around 47.97 to 53.05 dB and SSIM around 0.97 to 0.99, while also demonstrating significant adversarial robustness with elevated attack success rates against leading classifiers. Comparative and ablation studies indicate improved performance relative to representative watermarking, RDH-only, and adversarial-only approaches under the evaluated settings, suggesting that the framework is a practical and scalable solution for digital media protection.