Steganography enables covert communication by concealing information within innocuous media. Recent advances in deep learning have opened new avenues for high-capacity and imperceptible image steganography. This paper presents an experimental evaluation of a deep learning–based image-in- image steganography system employing an Encoder–Decoder– Discriminator architecture. The proposed framework embeds a secret image into a cover image to generate a visually indistinguishable stego image while enabling reliable recovery of the hidden content. The system is trained on a dataset of natural images using a composite loss function combining reconstruction, perceptual, structural similarity, and adversarial objectives. Quantitative evaluation using PSNR and SSIM metrics demonstrates strong imperceptibility of stego images and reasonable recovery fidelity of secret images. The results validate the feasibility of GAN-based approaches for practical image steganography and provide insights into their performance trade-offs.