保护云端监控视频隐私的密文内容检测技术研究

Video surveillance can exploit cloud for its storage and processing power and access from anywhere with any devices. In such a system, surveillance cameras simply stream video data to cloud, which then stores the video data, performs motion detection and other analyses, and alerts users if needed. Cloud video surveillance presents a unique challenge: privacy protection of surveillance video since it will be transmitted over public networks and stored at a third-party cloud that may not be trusted. This proposal addresses this challenging by proposing methods that allow a third party such as cloud to perform motion detection and other required analyses directly on encrypted and compressed surveillance video without decryption. In this way, cloud can still perform required detection yet the privacy of surveillance video is protected. In this proposal, we’ll investigate a whole spectrum of detection methods and corresponding encryption schemes on H.264 and HEVC compressed video to meet different needs of applications in content detection and privacy protection, from motion detection and trajectory detection and recognition, motion detection only without leaking any trajectory information, parameterized control on information leaking of encrypted surveillance video to meet different applications’ balances between privacy protection and video analytics in video surveillance, to a quantitative model to evaluate information leakage level of different encryption schemes on H.264 and HEVC compressed surveillance video. This proposed project will address both theoretic and practical issues in cloud video surveillance. We hope to attract more interests from relevant researchers and practitioners to this exciting emerging field to promote research and deployment of cloud video surveillance.

云视频监控可利用云端的存储和计算能力，并支持异构终端随时随地访问。监控摄像头仅需将视频数据传至云端存储，云端进行运动检测或分析，必要时发送告警信息。云视频监控面临特殊挑战：数据在公共网络传输和在第三方不可信云端存储带来隐私泄漏问题。为此本课题提出支持密文内容检测的隐私保护方法，允许第三方（如云端）在加密视频上直接进行运动检测或分析而无需解密，即可在保护隐私的同时满足云端的内容检测需求。本课题将针对H.264和HEVC压缩视频，探索一系列密文检测方案和对应加密方案，以满足不同应用场景对隐私保护和内容检测的多样化需求——从运动对象形状轨迹的检测与识别、保护形状轨迹的运动存在性判定，到隐私泄漏程度参数可控以应对动态变化的隐私保护和视频分析需求，并探索评估不同加密方案隐私泄漏程度的量化评价模型。本研究将吸引相关领域研究人员与业内人士加入对这一激动人心的新兴领域的研究，推动云视频监控的发展和普及。

大数据时代，智能终端在日常生活中的普及产生了海量数据，其所记录的数字内容也渐趋私人化。一方面，需要对这些海量数据进行智能化处理以充分利用其价值；另一方面，集中到云端的海量数据存在安全和隐私泄漏威胁，需要加以保护。本项目针对多媒体大数据应用，探索其智能处理的安全隐私保护关键技术。首先，考虑隐私受保护的云监控视频运动检测需求，提出了支持密文内容检测的隐私保护方法，在保护整个视频画面的同时允许第三方(如云端)在加密视频上直接进行运动检测或分析而无需解密，即可在保护隐私的同时满足云端的内容检测需求。本项目考虑不同应用场景对隐私保护和内容检测的多样化需求分别设计了不同隐私泄漏程度的方案并最终实现隐私泄漏程度参数可控。在此基础之上，我们进一步考虑在隐私保护前提下多媒体数据的密文信号处理需求，考虑多媒体数据在面对支持密文运算同态加密时所面临的密文膨胀问题带来的存储与传输开销巨大挑战，而多媒体数据解压缩过程上下文紧密相关，难以在同态域直接实现，我们重写解压缩算法，在无需解密的前提下，在同态加密的JPEG图片和FLAC音频密文上实现数据解压。