基于轻量级安全多方计算的分布式隐私数据合成OA
Lightweight Secure Multi-Party Computation for Distributed Private Data Synthesis
在大数据时代,跨域数据共享成为发掘数据价值的关键,但分散于多实体的高价值数据面临隐私泄露风险.现有分布式隐私数据合成方案通过安全多方计算协议实现差分隐私,但仅能处理低维数据集且计算与通信开销过高.为此,本文提出一种基于轻量级安全多方计算的差分隐私数据合成框架,通过分离隐私机制与后处理流程实现高效数据合成:安全计算节点基于安全多方计算协议在原始数据的统计信息上添加噪音,而服务器在明文环境下根据噪音结果生成合成数据集,缓解当前方案的计算瓶颈.基于此框架,本文提出分布式差分隐私数据合成方案 DisABSyn,其核心创新包括:重构原型方案的计算路径,将需要访问原始数据的高开销步骤,通过安全计算节点与服务器协同计算,减少高开销密文向量乘法;针对纵向场景设计基于边缘划分的隐私预算分配策略,提升该场景的数据效用.实验表明,DisABSyn 在横向与纵向场景的合成数据效用与原型方案相当.相比于现有的分布式数据合成方案,DisABSyn 方案保持更高数据效用并显著降低了安全多方计算的通信开销:在横向场景和纵向场景中,通信开销降低 88%和 99%.
In the era of big data,cross-entity data sharing has become crucial for exploring data value,while high-value data distributed across multiple entities faces privacy leakage risks.Existing dis-tributed privacy-preserving data synthesis schemes achieve differential privacy through high-overhead secure multi-party computation(MPC)protocols,which can only handle low-dimensional datasets and suffer from excessive computational and communication overhead.To address this issue,this study proposes a lightweight MPC-based differentially private data synthesis framework that separates pri-vacy mechanisms from post-processing,enabling efficient data synthesis:secure computing parties add noise to the statistical information of raw dataset via MPC protocols,while the server generates synthetic datasets in plaintext based on the noised results,relieving the computational bottlenecks of current schemes.Building on this framework,the distributed differentially private data synthe-sis scheme DisABSyn is presented,with core innovations including:restructuring the computational pipeline of the prototype scheme by replacing high-overhead steps requiring raw dataset access with collaborative computations between secure computing parties and the server,thereby reducing high-overhead ciphertext vector multiplications;a marginal partitioning-based privacy budget allocation strategy for vertical scenarios to enhance data utility in such settings.Experiments demonstrate that DisABSyn achieves synthetic data utility comparable to the prototype scheme in both horizontal and vertical scenarios.Compared to existing distributed data synthesis schemes,DisABSyn maintains higher data utility while significantly reducing the communication overhead of MPC,achieving an average reduction of 88%in horizontal scenarios and 99%in vertical scenarios.
贾靖宇;李鑫豪;刘哲理
南开大学 计算机学院,天津 300350||数据与智能系统安全教育部重点实验室(南开大学),天津 300350南开大学 密码与网络空间安全学院,天津 300350||数据与智能系统安全教育部重点实验室(南开大学),天津 300350南开大学 计算机学院,天津 300350||南开大学 密码与网络空间安全学院,天津 300350||数据与智能系统安全教育部重点实验室(南开大学),天津 300350
信息技术与安全科学
差分隐私数据合成安全多方计算
differential privacydata synthesissecure multi-party computing
《密码学报(中英文)》 2026 (4)
672-688,17
国家自然科学基金重点项目(62032012)Key Program of National Natural Science Foundation of China(62032012)
评论