基于模分量的同态加密方法研究与应用CSCD

随着云计算的发展,海量数据的处理正逐渐从用户本地转向云服务器,然而数据本身可能携带大量用户隐私,且一旦用户将数据上传至云服务器,就失去了对数据的完全掌控能力,该类数据一旦被非法获取,用户身份、行为、偏好等各类隐私就可能被暴露。因此,如何保证在不暴露原始数据的情况下让受委托的云服务器在密文下执行运算成为一个重要的研究课题。本文基于密码学和计算机视觉相关理论,针对隐私数据安全处理的问题,以模分量的同态性质为基础设计了两种加密方法,分别为基于混淆模分解的同态加密方法和基于密模聚合的同态加密方法,并给出了安全性分析。并将这两种方法应用于视觉盲计算领域中,实现计算方在无需获取任何原始数据有效信息的密文条件下,完成对数据的盲处理,实现了数据的可用不可见。实验结果表明,基于密模聚合模同态加密的运动目标盲提取方法,在多数测试场景中能在不降低原始算法准确率的前提下,在时间效率上明显优于基于混合高斯模型的运动目标盲提取和基于多服务器秘密共享的前景提取等方法;基于混淆模分解同态加密的人脸盲检测方法,能在不降低原始人脸检测算法识别的准确率前提下,实现视频监控人脸的盲检测,且检测速度大幅度快于基于随机子图的隐秘人脸检测方法和基于随机向量的隐秘人脸检测等算法。

Research and Implementation of Blind Video Processing Method Based on Modular Component HomomorphismCSCD

With the development of cloud computing, the processing of massive data is gradually shifting from local users to cloud servers. However, the data itself may carry a lot of user privacy, and once users upload the data to the cloud servers, they will lose their full control over the data. Once such data is illegally obtained, all kinds of privacy such as user identity, behavior and preferences may be exposed, and the consequences will be unimaginable. Therefore, it has become an important research topic how to ensure that the entrusted cloud server can perform operations under ciphertext without exposing the original data. Based on the related theories of cryptography and computer vision, aiming at the problem of privacy data security, this paper designs two encryption methods based on the homomorphism of modular components, namely, homomorphism encryption method based on confusing modular decomposition and homomorphism encryption method based on dense modular aggregation, and gives a detailed security analysis. These two methods are applied to the field of visual blind computing, so that the computing side can complete the blind processing of the data without obtaining any ciphertext of the original data, and the data is available and invisible. The experimental results show that the blind extraction method of moving target based on dense mode aggregation homomorphism encryption is obviously superior to blind extraction of moving target based on mixed Gaussian model and foreground extraction based on multi-server secret sharing in most test scenarios without reducing the accuracy of the original algorithm. The blind face detection method based on homomorphic encryption of Confused Modulus Decomposition can realize blind face detection in video surveillance without reducing the recognition accuracy of the original face detection algorithm, and the detection speed is greatly faster than the hidden face detection method based on random subgraph and the hidden face detection algorithm based on random vector. © 2023 Chinese Academy of Sciences. All rights reserved.