基于联邦学习的本地化差分隐私机制研究

联邦学习与群体学习作为当前热门的分布式机器学习范式，前者能够保护用户数据不被第三方获得的前提下在服务器中实现模型参数共享计算，后者在无中心服务器的前提下利用区块链技术实现所有用户同等地聚合模型参数。但是，通过分析模型训练后的参数，如深度神经网络训练的权值，仍然可能泄露用户的隐私信息。目前，在联邦学习下运用本地化差分隐私(LDP)保护模型参数的方法层出不穷，但皆难以在较小的隐私预算和用户数量下缩小模型测试精度差。针对此问题，该文提出正负分段机制(PNPM)，在聚合前对本地模型参数进行扰动。首先，证明了该机制满足严格的差分隐私定义，保证了算法的隐私性；其次分析了该机制能够在较少的用户数量下保证模型的精度，保证了机制的有效性；最后，在3种主流图像分类数据集上与其他最先进的方法在模型准确性、隐私保护方面进行了比较，表现出了较好的性能。     

基于生成对抗网络的差分隐私数据发布方法

机器学习的飞速发展使其成为数据挖掘领域最有效的工具之一,但算法的训练过程往往需要大量的用户数据,给用户带来了极大的隐私泄漏风险.由于数据统计特征的复杂性及语义丰富性,传统隐私数据发布方法往往需要对原始数据进行过度清洗,导致数据可用性低而难以再适用于数据挖掘任务.为此,提出了一种基于生成对抗网络（Generative Adversarial Network,GAN）的差分隐私数据发布方法,通过在GAN模型训练的梯度上添加精心设计的噪声来实现差分隐私,确保GAN可无限量生成符合源数据统计特性且不泄露隐私的合成数据.针对现有同类方法合成数据质量低、模型收敛缓慢等问题,设计多种优化策略来灵活调整隐私预算分配并减小总体噪声规模,同时从理论上证明了合成数据严格满足差分隐私特性.在公开数据集上与现有方法进行实验对比,结果表明本方法能够更高效地生成质量更高的隐私保护数据,适用于多种数据分析任务.     

基于差分隐私的轨迹隐私保护方法

针对现有的轨迹隐私保护模型大多难以抵御复杂背景知识攻击的问题,本文提出了一种基于差分隐私的轨迹隐私保护方法.首先结合地理不可区分机制对原始轨迹数据添加半径受限的拉普拉斯噪音;其次构造数据映射模型将原始数据和噪音数据映射到新的发布位置,使攻击者无法获取真实轨迹数据;接着应用最优数据映射函数发布最优的轨迹位置以提高发布数据的可用性;最后利用差分隐私抵御非敏感信息推理攻击,进一步保护用户隐私.实验结果表明,本文算法既能有效保护轨迹数据中用户的隐私,也能保证数据的可用性.     

差分隐私技术在科技业务平台用户数据保护中的研究

随着部门科技业务平台与外部省级平台的接入,系统间的数据请求将不可避免引起用户数据泄露问题。部门平台改造升级过程中,虽然通过国产密码杂凑算法SM3构建用户认证体系,同样面临差分攻击的风险。针对平台用户隐私泄露、数据可用性等问题,研究引入差分隐私保护技术来应对平台数据请求、共享发布过程中带来的用户敏感数据泄露的风险。同时对差分隐私、数据集算法组合、噪声机制进行了概要叙述,描述了差分隐私在科技业务平台非交互式数据发布中的应用场景。     

一种新的满足隐私性的云存储公共审计方案

在云存储网络环境中,数据的安全性、完整性和隐私性是用户最关心的问题之一.云存储服务中,用户将存储的数据和认证标识信息存储在云服务器中.为了保证存储数据的完整性,云存储服务提供者需要向用户或第三方审计者证明其正确地持有用户存储的数据.公共审计是指由用户以外的第三方代替用户完成审计工作,这对于计算资源比较有限的用户尤其重要.目前多数云存储审计方案没有考虑隐私性问题.本文提出了一种新的可聚合基于签名的广播加密(ASBB)方案,并在此基础上设计了新的满足隐私性的云存储公共审计方案.新方案在随机预言模型下是可证安全的,并且在计算开销方面更具有优势.     

基于深度学习的位置大数据统计发布与隐私保护方法

针对传统位置大数据统计划分发布结构不合理、划分发布方法效率低下的问题,提出一种基于深度学习的位置大数据统计划分结构预测方法和差分隐私发布方法,以提高位置大数据统计划分发布数据的可用性和执行效率.首先对二维空间进行细致划分和自底向上合并,从而构建合理的空间划分结构.然后将划分结构矩阵组织为三维时空序列,借助深度学习模型提...     

异构社交网络中用户轨迹隐匿方法优化研究北大核心CSCD

在对异构社交网络中用户轨迹进行隐匿时,当前方法大多针对用户单个位置进行轨迹隐匿,不适于复杂的异构社交网络.为此,提出一种新的基于假轨迹的异构社交网络中用户轨迹隐匿方法,通过一个例子对所提方法的基本思想进行分析.对假轨迹方法进行概述,在中心服务器系统结构上实现.给出单个位置的暴露风险、轨迹暴露风险、距离偏移度的概念和计算公式.为了避免攻击者判断出用户真实轨迹的概率,提出记忆规则进行优化.将查询消息、真实位置等参数传输至隐私保护服务器,隐私保护服务器依据用户参数要求和记忆规则产生满足条件的假位置,将含有假位置的匿名框传输至服务提供商,隐私保护服务器对返回的查询结果进行求精后传输至用户.实验结果表明,采用所提方法得到的轨迹数据有很高的可用性.     

基于差分隐私的活动模式保护与时空轨迹发布方法

为了解决用户轨迹数据发布时的活动模式泄露问题，本文提出了一种基于差分隐私的活动模式保护与时空数据发布方法 DPAP-STTP(Differentially Private Activity Pattern and Spatial-Temporal Trajectory Publication)，该方法即保护了用户时空数据中活动模式的隐私，又可以保证所发布时空轨迹在服务建议生成上的有效性.在DPAP-STTP中，用户的活动模式表示为个人代表性轨迹的动静态信息，包括代表性轨迹的时空密度分布、时空路径分布、移动模式以及时空跨度.另外，DPAP-STTP通过隐私保护预算与隐私保护阈值对该动静态信息进行调控，然后根据调控后的动静态信息依次划分时空网格、重构轨迹所处时空区间、时空轨迹点随机采样，最终生成满足群体差分隐私的时空轨迹进行发布.本文的实验比较了DPAP-STTP与DP-STAR(Differential Private Synthetic Trajectory Publisher)、BNA(Bounded Noise-Adding)所生成的轨迹在特定时空范围内的有效性，证明DPAP-ST...     

支持第三方仲裁的智能电网数据安全聚合方案

智能电网作为新一代的电力系统,显著提高了电力服务的效率、可靠性和可持续性,但用户侧信息安全问题也日渐突出.本文针对智能电网系统中用户数据信息泄露的问题,提出了一个具有隐私保护的数据安全采集方案.收集器能够对采集到的电表数据进行验证,聚合为一个新的数据包,发送给电力服务中心解密和存储,且第三方仲裁机构能够解决用户端智能电表与电力服务中心发生的纠纷.同时,本方案支持收集器,电力服务中心和第三方仲裁机构执行批量验证操作,以提升验证效率.本文的理论分析与实验比较表明,该方案比同类型方案具有更高的运算效率和通信效率.     

基于差分隐私模型的位置轨迹发布技术研究

位置轨迹大数据的安全分享、发布需求离不开位置轨迹隐私保护技术支持。在差分隐私出现之前,K-匿名及其衍生模型为位置轨迹隐私保护提供了一种量化评估的手段,但其安全性严重依赖于攻击者所掌握的背景知识,当有新的攻击出现时模型无法提供完善的隐私保护。差分隐私技术的出现有效地弥补了上述问题,越来越多地应用于轨迹数据隐私发布领域中。该文对基于差分隐私理论的轨迹隐私保护技术进行了研究与分析,重点介绍了差分隐私模型下位置直方图、轨迹直方图等空间统计数据发布方法,差分隐私模型下轨迹数据集发布方法,以及连续轨迹实时发布隐私保护模型。与此同时,在对现有方法对比分析的基础上,提出了未来的重点发展方向。     

Location publishing technology based on differential privacy-preserving for big data services

Aiming at dealing with prospect knowledge and complex combinatorial attack,a new location big data publishing mechanism under differential privacy technology was given.And innovative usability evaluation feedback mechanism was designed.It gave corresponding solution details for the sensitive attributes and the identity recognition to analyze the quality of service,aimed at privacy protecting for location based big data under situations like combination of location information and non-location information and attacker’s arbitrary background knowledge.Simulation results based on different spatial indexing technology proved that the new publishing model has a higher accuracy under specified privacy conditions for the location query service.     

FedCDR: Privacy-preserving federated cross-domain recommendation

Cross-Domain Recommendation (CDR) aims to solve data sparsity and cold-start problems by utilizing a relatively information-rich source domain to improve the recommendation performance of the data-sparse target domain. However, most existing approaches rely on the assumption of centralized storage of user data, which undoubtedly poses a significant risk of user privacy leakage because user data are highly privacy-sensitive. To this end, we propose a privacy-preserving Federated framework for Cross-Domain Recommendation, called FedCDR. In our method, to avoid leakage of user privacy, a general recommendation model is trained on each user's personal device to obtain embeddings of users and items, and each client uploads weights to the central server. The central server then aggregates the weights and distributes them to each client for updating. Furthermore, because the weights implicitly contain private information about the user, local differential privacy is adopted for the gradients before uploading them to the server for better protection of user privacy. To distill the relationship of user embedding between two domains, an embedding transformation mechanism is used on the server side to learn the cross-domain embedding transformation model. Extensive experiments on real-world datasets demonstrate that our method achieves performance comparable with that of existing data-centralized methods and effectively protects user privacy.     

Approach of location privacy protection based on order preserving encryption of the grid

The centralized structure of the trusted third party is a major privacy protection structure on location based services.However,if the central third party server can not be trusted or compromised,users have the risk of leakage of privacy location.Aiming at the above problems,location privacy protection approach based on a user-defined grid to hide location was proposed.The system first automatically converted the query area into a user-defined grid,and then the approach utilized order preserving encryption,which made the user’s real-time position in the hidden state could still be compared.Because the information in the process of the approach was in a state of encryption,the server could not know the user’s location information,thus improved privacy protection of the user location.The central third party server only need to do simple comparison work,so its processing time overhead would effectively decrease.Security analysis certificate the security of the proposed approach and simulation experimental show the proposed approach can reduce the time cost of the central third party server.     

Research on LBS privacy preservation based on pseudorandom permutation in road network

A method of privacy preservation based on pseudorandom permutation was put forward for the issues of location privacy and query content privacy.Firstly,the distribution information of points of interest (PoI) based on the vertexes in the road network was organized,each single road vertex was taken as the foundational processing object.Based on the pseudorandom permutation,a permutation scheme of the point-of-interest records at the LBS server's end was put forward,a 32-bit random seed was adopted to generate a permuted table in the scheme,and the point-of-interest records were encrypted and permuted according to the table.These processed records were stored in the LBS database.Then a trusted intermediate server,replacing of the user,issued a query request with a record number instead of the query content to the LBS server.The LBS server could not determine which kind of PoI the user was interested in or which road section the user was locating on,and therefore the scheme achieved private information retrieval.Finally,the efficiency in the metrics of query accuracy,communication overhead and processing time was also analyzed.By the performance analysis and extensive experiments,the proposed scheme is proved to be location untraceable and query content uncorrelation.     

CLM:differential privacy protection method for trajectory publishing

In order to solve the problem existing in differential privacy preserving publishing methods that the independent noise was easy to be filtered out,a differential privacy publishing method for trajectory data (CLM),was proposed.A correlated Laplace mechanism was presented by CLM,which let Gauss noises pass through a specific filter to produce noise whose auto-correlation function was similar with original trajectory series.Then the correlated noise was added to the original track and the perturbed track was released.The experimental results show that the proposed method can achieve higher privacy protection and guarantee better data utility compared with existing differential privacy preserving publishing methods for trajectory data.     

Survey on differential privacy and its progress

With the arrival of the era of big data sharing,data privacy protection issues will be highlighted.Since its introduction in 2006,differential privacy technology has been widely researched in data mining and data publishing.In recent years,Google,Apple and other companies have introduced differential privacy technology into the latest products,and differential privacy technology has become the focus of academia and industry again.Firstly,the traditional centralized model of differential privacy was summarized,from the perspective of analysis of data mining and data released in the differential privacy way.Then the latest local differential privacy regarding data collection and data analysis based on the local model was described,involving crowdsourcing with random response technology,BloomFilter,statistical inference techniques.Finally,the main problems and solutions of differential privacy technology were summarized.     

Method of privacy protection based on multiple edge servers in personalized search

In the plaintext environment,users' personalized search results can be obtained through users' interest model and query keywords.However,it may possibly result in the disclosure of sensitive data and privacy,which prevents using sensitive data in cloud search.Therefore,data is generally stored in the form of ciphertext in the cloud server.In the process of cloud search service,users intend to quickly obtain the desired search results from the vast amount of ciphertext.In order to solve the problem,it was proposed that a method of privacy protection based on multiple edge servers in personalized search shall be used.By introducing multiple edge servers and cutting the index as well as the query matrix,the computing relevance scores of partial query and partial file index are achieved on the edge server.The cloud server only needs to get the relevance score on the edge server and make a simple processing that can return to the most relevant Top K files by user query,so as to make it particularly suitable for a large number of users in the massive personalized ciphertext search.Security analysis and experimental results show that this method can effectively protect users’ privacy and data confidentiality.In addition,it can guarantee high efficiency in search to provide better personalized search experience.