The lack of effective copyright protection for AI-generated artworks in style transfer remains a significant challenge. Existing approaches typically embed watermarks as post-processing steps after image generation, separating copyright protection from the artistic creation process. This leads to shallow embedding that degrades style quality. To address this limitation, this paper proposes Perceptual and Copyright Dual Generative Adversarial Networks (PCD-GAN), a novel framework that seamlessly integrates copyright protection into the artistic creation process. Specifically, a copyright-modulated Adaptive Instance Normalization (CM-AdaIN) mechanism embeds copyright information by modulating normalization parameters, preserving style statistics while achieving deep semantic coupling. In addition, a frequency-domain copyright injection module embeds copyright signals into the mid-frequency subbands of the discrete wavelet transform (DWT) using adaptive strength and pseudo-random spatial expansion, providing pixel-level robustness that complements feature-level embedding. Furthermore, a dual-discriminator collaborative adversarial training architecture employs separate discriminators for visual authenticity and copyright correctness. This design decouples perceptual quality and copyright objectives, effectively reducing gradient conflicts during training. Experimental results demonstrate state-of-the-art performance, achieving a PSNR of 28.56 dB and a copyright extraction accuracy of 93.5% under various attacks. Compared with existing methods, the proposed approach improves robustness by 8.1% under JPEG compression while maintaining imperceptible watermark embedding.
Lyu et al. (Thu,) studied this question.