面向网络视频直播的敏感信息实时检测技术研究将为减少网上不良信息传播、维护国家安全和社会稳定提供重要技术支撑。研究融合弹幕和用户观看行为的快速敏感直播频道过滤方法；研究基于弹幕、用户行为和语音联合特征的视频直播敏感信息实时分析技术；研究基于深度学习技术的敏感直播视频/图像实时特征提取技术。通过对网络视频直播平台中弹幕、音视频多媒体数据和打赏互动关系的系统分析，研究多元信息的敏感内容检测特征的有效组合遴选方法。融合弹幕、音视频多媒体数据的低层特征和主播互动行为特征，并结合主播的直播历史画像，研究文本、音视频的底层特征与敏感视频信息之间的映射关系；研究基于多元特征融合的视频直播敏感信息实时检测方法。将重点探索实时/近实时深度学习框架以提高视频直播中敏感信息检测的实时性能，以及融合直播内容低层数据特征获取高层语义的机理和模型，为我国网络视频直播监管提供新的理论基础和核心算法。

The investigation of sensitive information detection in real-time for Internet video live broadcast plays a vital role in terms of reducing the spread of harmful information, safeguarding national security and social stability. In our research project, we will study the following major aspects: rapid filters of sensitive live channels by combining the features of bullet screens and user behaviors, real-time analysis of sensitive information by analyzing bullet screens, user behaviors and audio data, deep learning-based analysis of sensitive videos and images in the live streaming, and multi-feature fusion-based detection of sensitive information in the video live streaming. We will find effective features for sensitive information detection by analyzing the sensitive content of bullet screens and audio data, and user behaviors in sensitive live channels. Besides, we will build a multi-feature fusion-based multi-objective hierarchical detection framework to realize the fast detection of sensitive contents by leveraging the user profile of an anchor. This project will focus on two scientific challenges, which could provide the theoretical bases and the core algorithms for the relevant innovative applications: the first one is to research the mechanism of obtaining high-level semantic information from low-level fusion features of the Internet live streaming, and the other one is to study the real-time/near-real-time deep learning framework for sensitive information detection from Internet video live broadcast.