Animated GIF has become a key communication tool in contemporary social platforms thanks to highly compatible with affective performance, and it is gradually adopted in commercial applications. Therefore, the copyright protection of the animated GIF requires more attention. Digital watermarking is an effective method to embed invisible data into a digital medium that can identify the creator or authorized users. However, few works have been devoted to robust watermarking for the animated GIF. One of the main challenges is that the animated image also contains time frame dimension information compare with still images. This paper proposes a robust blind watermarking framework based 3D convolutional neural networks for the animated GIF image, which achieves watermark image embedding and extraction for the animated GIF. Also, noise simulation is developed in frame-level to ensure robustness for the attack of the temporal dimension in this framework. Furthermore, the invisibility of the watermarked animated image is optimized by adversarial learning. Experimental results provide the effectiveness of the proposed framework and show advantages over existing works.

GIF动画由于与情感表现的高度兼容，已成为当代社交平台的重要交流工具，并逐渐被商业应用所采用。因此，动画GIF的版权保护需要更多的关注。数字水印是一种将不可见数据嵌入到可以识别创建者或授权用户的数字媒体中的有效方法。然而，很少有工作致力于动画 GIF 的鲁棒水印。主要挑战之一是与静止图像相比，动画图像还包含时间帧维度信息。本文提出了一种基于3D卷积神经网络的鲁棒盲水印框架，用于动画GIF图像，实现了动画GIF的水印图像嵌入和提取。还，噪声模拟是在帧级开发的，以确保该框架中时间维度攻击的鲁棒性。此外，水印动画图像的不可见性通过对抗性学习进行了优化。实验结果提供了所提出框架的有效性，并显示出优于现有工作的优势。