云社交平台中内容感知服务的隐私保护研究

With the proliferation of camera-equipped mobile devices, images are becoming one of the key enablers of users’ connectivity at social media sites. For the semantic richness of image content, many social applications are directly utilizing images for various social discovery purposes. To handle the storage and computation involved with the explosive increasing of images, many social media sites opt to leverage the commercial public cloud as their backend datacenter, where users’ uploaded images go straight to cloud rather than passing through their on-premise physical infrastructures. While extremely convenient, building image based social discovery over this outsourced architecture also introduces acute privacy concerns. First, images are extremely easy to reveal content-sensitive information, such as aspects of users’ private sphere, affecting both the subjects in the content and the content owner. Besides, any potential unnecessary leakage from image-based social discovery may cause higher privacy risks, as it simply exposes personal profiles, such as background, living places, family members, etc., to the public cloud, where certain new security threats are still not yet fully understood. Given that many images now directly go to the public cloud in unencrypted form, various internal/external security attacks at cloud make the potential threats of user privacy leakage even more intimidating..To address these challenges, we propose to investigate and prototype a privacy-preserving content-centric social discovery system. It aims to provide secure, efficient, and accurate user recommendations directly over encrypted image content, without revealing unnecessary information to public clouds. Our first research module is to explore a suite of techniques to enable encrypted user profile indexing, which supports the core of image-based social discovery via fast and scalable similarity search over millions of encrypted user profile vectors. Then we propose to investigate a more complicated scenario where user/data dynamics are to be supported without affecting the correctness and security of the design. We plan to conduct thorough security analysis under well-defined security models. We also plan to perform comprehensive system evaluation based on large-scale of image datasets so as to fine tune the best combination of system parameters for various application contexts. Our proposed research is expected to address the privacy risks for users who are increasingly using images for online social interactions in the cloud era. It will serve as a steppingstone to more advanced privacy-aware social applications built on top of users’ encrypted content, and naturally complements with existing context-based approaches. Our research methodologies will extend the scientific frontier from security, information retrieval, and mobile social networks. We will use results from this project as educational tools for undergraduate/graduate students research training.

随着多媒体技术的飞速发展和智能移动设备的普及，多媒体（如图片、视频）已成为社交网络中关联用户的重要媒介。典型地，图片具有内容形象、语义丰富等特性，这促使社交网络通过图像内容感知技术来推荐好友。为应对数据的爆炸性增长，社交网络逐渐利用商业云平台作为其数据存储后端。但是，将内容感知相关的服务外包到公有云会带来严重的安全隐患。本项目致力于研究并实现一个内容感知并保护隐私的社交服务推荐系统，立足于安全、高效、准确地在加密图像上进行好友推荐，从而使敏感信息不被泄露。首先，研究构建经加密的用户档案索引，并支持加密相似搜索。其次，研究用户数据的安全更新方法，保证其正确性和安全性。计划在严格定义的安全模型下详细论证系统的安全性。最后，对系统进行全面的性能和效果分析。本项目旨在解决云外包背景下，用户使用图片进行社交活动时带来的隐私泄露问题。对于其他隐私内容感知的社交应用来说，本项目的成果将会是一个奠基石。

随着多媒体技术的飞速发展和智能移动设备的普及，多媒体（如图片、视频）已成为社交网络中关联用户的重要媒介。为应对多媒体数据的爆炸性增长，社交网络中产生的数据逐渐改为交由公有云代为存储。然而，由此造成的安全隐患已成为阻碍社交网络发展的桎梏。针对此现状，本项目着重研究和设计一个支持隐私保护的图像内容感知的好友推荐系统。研究内容包括：1.针对加密数据的相似性搜索技术、确定性搜索技术以及范围搜索技术；2.针对加密视频数据的网络内去重技术；3.安全的云外包中间件服务；4.支持隐私保护的行为定向技术；5.安全的传感数据真值学习技术；6.面向云社交多媒体服务的隐私保护图像去噪、重复删除和共享技术；7.防欺骗的语音认证系统。本项目在CCS 2019、INFOCOM 2019、INFOCOM 2016、ICDCS 2018、ASIACCS 2017、IEEE TPDS、IEEE TDSC、IEEE TIFS等著名国际会议和国际期刊发表/录用了37篇论文其中SCI检索19篇，EI检索18篇。本项目培养了7位博士生，申请了专利5件，很好地完成了项目既定的研究目标。取得了如下主要成果：1.提出了一种隐私保护的相似性搜索方案，隐藏了搜索频率且支持结果共享，方案部署在云端实现了对加密高维数据的高效搜索；2.提出了一种基于网络内缓存的高效加密视频传输协议，支持自适应视频传递和访问控制，严格的分析和原型评估证明了设计的安全性和高效性；3.设计了安全的网络外包中间件服务，支持对加密数据包的动态检测，系统部署在真实的云外包网络中间件上展现出实用性；4.提出了支持隐私保护的行为定向方法，满足了实际应用中丰富的功能需求；5.提出了多种基于多媒体数据的隐私保护应用技术，包括去噪、重复删除、共享和社群发现，当在真实的数据集上进行测试时，这些技术均展现出其高效性；6.设计了一款智能手机防欺骗语音系统，系统部署在智能手机上实现了高准确率；7.提出了支持隐私保护的多媒体传感数据真值学习技术，提升了云社交中内容感知服务的数据价值。本项目的研究成果为支持隐私保护的多媒体数据内容感知服务提供技术支持，为计算外包背景下的社交网络应用的开发与推广提供重要依据。