A game-theoretic deep learning framework for secure image encryption and steganography

Secure image transmission remains an important issue in multimedia security, especially when adversaries can detect or analyze hidden communication. This study proposes an empirical game-theoretic deep learning framework that integrates image steganography with a key-conditioned latent protection mechanism. The framework consists of an encoder, a decoder, and an adversarial detector; independent steganalysis detectors are used only for held-out evaluation. The encoder embeds a key-conditioned latent representation of the secret image into a cover image, the decoder reconstructs the secret image with the matched key, and the adversarial detector estimates whether an image is cover or stego. To improve reproducibility, the revised setting specifies the players, strategy spaces, utilities, threat model, key protocol, losses, dataset splits, payload definition, and metric calculation. On the pooled BOSSBase 1.01 and ImageNet held-out test sets, the proposed method reports cover-to-stego PSNR of 40.13 dB, cover-to-stego SSIM of 0.978, secret bitwise BER of 1.92%, internal adversary accuracy of 57.2%, independent-detector accuracy of 60.8%-64.7%, AUC of 0.604-0.671, and measured embedding capacity of 4.71 bpp under the stated operating protocol. These results indicate reduced, but not eliminated, empirical detectability. The security claims are therefore limited to empirical resistance under the specified threat model and should not be interpreted as formal cryptographic security. Not applicable.