Abstract: This study addresses the issues of low computational efficiency, excessive parameter quantities, and insufficient accuracy in discriminating watermark authenticity observed in existing image watermark detection algorithms. An improved GAN-based image watermark detection algorithm is proposed.The algorithm comprises two models, DSD-GAN and DCG-GAN, which are designed for watermark presence detection and watermark authenticity discrimination, respectively.The discriminator of the DSD-GAN model is optimized by incorporating depthwise separable convolution and dilated convolution, where the parameter quantity is effectively reduced and computational efficiency is enhanced. Significant improvements in both speed and accuracy of watermark presence detection are achieved through this architectural refinement.The DCG-GAN model integrates a generator where high-quality forged watermarks are synthesized, and the discriminator parameters are optimized via the Grey Wolf Optimization (GWO) algorithm, leading to a significant improvement in the accuracy of watermark authenticity discrimination.The watermark presence detection and authenticity discrimination are efficiently achieved through the collaborative workflow of the two models, where task-specific feature extraction and adversarial optimization are seamlessly integrated.Experimental results demonstrate that watermark imperceptibility and robustness are ensured while a detection accuracy of 99.58% is achieved by the DSD-GAN model in watermark presence detection tasks, indicating superior performance. For the DCG-GAN model in watermark authenticity discrimination, excellent PSNR and NC metrics are observed, and the accuracy is consistently maintained above 0.50 across varying embedding strengths, ensuring the imperceptibility and robustness of the watermark are comprehensively preserved.The parameter quantity is effectively reduced and computational efficiency is enhanced in the proposed solution compared to other algorithms, while a more efficient solution is provided for image watermark detection in complex scenarios through architectural optimization and lightweight design strategies.