基于特征哈希的云端海量医疗图像安全检索研究

Secure retrieval over large-scale medical images, is a kernel issue in secure sharing and privacy protection of health big data in cloud. In real cloud environment, existing schemes suffer from four major limitations, such as the security risk without considering the leakage of result pattern, the serious efficiency degradation caused by heavy computation burden, unbalanced computation load between cloud and client, and weak support for high-dimensional data. This project aims at achieving secure and efficient retrieval over large-scale medical images in cloud, with resolving the essential problem of approximate nearest neighbor search over encrypted image features. The research work will be done from the aspects of security model, approximate nearest neighbor search algorithm, and outsourcing mechanism. First, for hiding result pattern and improving the security of image retrieval, an appropriate security model will be investigated based on model combination and pattern hidden mechanism. Based on security, approximate nearest neighbor search schemes over encrypted image features will be proposed by levering locality sensitive hashing-based fine-grained secure division mechanisms, lightweight cryptographic tools, and universally secure similarity measurements, in order to overcome the curse of dimensionality and promote the efficiency of approximate nearest neighbor search. Furthermore, outsourcing mechanism of secure image retrieval will be studied, in which refinement mechanism is securely outsourced to a cloud, to achieve securely and rapidly retrieval and matching over large-scale medical image data. Also, an application demonstration will be constructed based on the above research work. Through the research work over this project, the achievements and application demonstrations will resolve the problems of privacy protection and efficiency bottleneck better than ever while pushing the constructions in application infrastructures of health big data sharing in cloud.

海量医疗图像的安全检索，是云端健康大数据安全共享与隐私保护的核心问题。现有方案在云环境中存在以下不足：忽略结果模式的安全模型易导致安全隐患、密码学方案的计算开销大导致检索效率严重退化、云端和客户端计算负载失衡、对高维数据支撑不足。本项目为实现云端海量医疗图像的安全高效检索与匹配，拟解决加密图像高维特征的近似最近邻查询这一关键问题，从安全模型、近似最近邻查询算法、外包处理机制三方面展开研究。研究图像检索的安全模型，基于模式隐藏、模型组合等技术，隐藏结果模式并增强图像检索的安全性；在安全保障的基础上，研究加密图像高维特征的近似最近邻查询算法，基于安全划分、普适安全度量机制，降低维数灾难影响并提升近似最近邻查询效率；进一步，研究安全图像检索外包处理机制，将安全精炼外包给云，实现海量医疗图像云端的安全高效检索、匹配及应用。通过研究，产生的成果将更好地解决健康大数据云端共享的隐私保护和效率瓶颈问题。

课题组在研究过程中，重点突破普适的高维数据近似性安全度量机制和安全图像检索中结果模式的隐藏机制两个关键科学问题，以加密图像高维特征的近似最近邻查询算法为技术手段，开展云端海量医疗图像安全检索的研究。研究构建面向图像数据检索的安全模型，充分考虑结果模式泄露和模型迁移中检索结果的近似性这两者对模型安全性的影响，结合图像特征的固有特性，构建适用于图像检索的自适应语义安全模型，为实现图像高效安全检索打下坚实基础。针对加密图像的高维特征，提出高维数据的查询点邻域安全定位技术和多种度量空间下的安全精炼技术，实现近似最近邻查询方案效率和安全性的同步提升。研究脱敏处理技术、加密图像特征近似最近邻查询的细粒度外包处理技术，搭建基于混合云的在线医疗图像安全检索与匹配应用示范平台，提高了云端共享医疗图像数据的安全性和可操作性。通过上述三个方面的研究，完成了海量医疗图像的云端安全、高效检索与匹配的项目既定总体目标。课题组通过三年的持续研究，共发表已进行资助标注的高水平学术论文10篇，其中在CCF A类刊物IEEE TDSC上取得了突破，并在CCF A类中文刊物《计算机学报》和《密码学报》上各发表1篇论文，SCI检索论文共计7篇。以第一发明人身份共申请国家发明专利4项，其中1项专利已经获得授权，完成了1项国家发明专利的转让。协助所在团队老师培养博士生1名，硕士生3名。全部考核指标均已完成，部分考核指标超出任务合同书要求。.综上，课题组按照既定计划完成了各项研究任务。