Chroma-Warp GAN: High-Capacity Reversible Image Steganography via Perceptually Guided Chrominance Modulation

Abstract

Deep learning-based image steganography enables secure covert communication by hiding information within images imperceptibly. Although recent GAN-based methods with hybrid attention improve stego image quality and secret recovery, they operate in RGB space, modifying all channels and causing visible distortions. Since the human visual system is more sensitive to luminance changes than to chrominance variations. To address this, Chroma-Warp GAN operates in the YUV colour space. The generator produces a two-channel secret embedding via a convolutional-residual pathway and fuses it with the cover UV channels, where a trainable UVScaleLayer applies a clipped adaptive scalar to generate a controlled chrominance residual, leaving the Y channel unchanged. The extractor recovers the secret via a sigmoid branch, reconstructs the cover through scaled residual addition, and predicts a UV consistency signal for auxiliary supervision. A discriminator with three strided convolutional layers and global average pooling enforces adversarial realism through a staged multi-loss objective. Experiments on the DOTA dataset achieve Cover–Stego PSNR of 43.26 dB, Secret–Recovered Secret PSNR of 42.35 dB, and Cover–Recovered PSNR of 42.74 dB, confirming effective chrominance-restricted embedding with reliable extraction and lossless reversibility.