基于深度学习的抗编辑视频水印技术

非法用户借助视频编辑工具盗用或未经授权篡改视频内容，不仅带来侵权问题，甚至诱发群体事件，如2018年短视频平台下架57万部作品、新疆“7.5”事件。盗版和伪造的视频严重影响市场秩序和地区和平稳定。现有视频认证技术对编辑视频内容是否合法的智能判断存在一定局限性，迫切需要有效的技术手段对编辑视频、伪造视频、偷摄视频的真实性进行认证。.本项目拟研究基于深度学习的抗编辑视频水印核心算法及评测方法，主要包括：融合时空信息的视频自适应特征选取方法、基于深度网络体系与感知原理的抗编辑的视频水印嵌入技术、基于卷积神经网络的视频水印提取技术。在此基础上，构建视频认证数据集，制定认证性能评测方法。本项目的研究成果将为公共视频真实性认证及应用提供相应的技术支撑。

Video content is pirated or edited by illegal users, which not only brings about infringement act, but also induces group incidents, such as the case that 570,000 video works of short video platform have been removed in 2018, the Xinjiang "7.5" incident, and so on. Videos pirated and forged will seriously affect market order, regional peace and stability. The existing technologies cannot intelligently judge the legality of videos’contents. It is urgent that researching innovative and effective technical methods to authenticatethe realness of editing videos, tampered videos and candid videos..This project intends to study some key algorithms and evaluation methods of anti-editing video watermark based on deep learning. The contents mainly include the methods of selecting self-adaptive video feature mixing spatial-temporal information, the anti-editing video watermark embedding techniques based on deep network system, perceptual principleand, the video watermark extraction techniques based on convolutional neural network, what's more, we will construct the video authentication dataset and formulate the authentication performance evaluating methods. The research results of this project will provide technical support for the realness authentication and applications of public videos.