MTGAN: research on video watermarking based on multi-task generative adversarial networks

In response to the shortcomings of traditional video watermarking in terms of invisibility and robustness, this paper proposes a video watermarking algorithm based on a multi-task generative adversarial network (MTGAN). By organically integrating the advantages of multi-task learning and generative adversarial networks, a generator with shared parameters and a multi-task discriminator are designed. This approach jointly optimizes the tasks of video generation and watermark information extraction, simultaneously achieving high visual quality (with both peak signal-to-noise ratio and structural similarity index reaching satisfactory levels) and significantly improving the recovery accuracy of watermark information (with the NC metric attaining a commendable level). The UCF-101 and STL-10 datasets are utilized to validate the model’s generalization capability. Compared with representative baselines, the proposed method achieves a favorable trade-off between imperceptibility and robustness, providing more stable visual quality while improving robustness under common spatial and codec-related distortions.