一种基于隐私保护下的多方记录链接方法

多方隐私保护下的记录链接（privacy-preserving record linkage，PPRL）是在隐私保护下从多个数据源中找出代表现实世界中同一实体的过程，该过程除了最终匹配结果被数据源之间共享，其他信息均未被泄露.随着数据量的日益增大和现实世界数据质量问题的存在（如拼写错误、顺序颠倒等），多方PPRL方法的可扩展性和容错性面临挑战.目前，已有的大部分多方PPRL方法都是精确匹配方法，不具有容错性.还有少部分多方PPRL近似方法具有容错性，但在处理存在质量问题的数据时，由于容错性差和时间代价过大，并不能有效地找出数据源间的共同实体.因此，本文提出一种结合布隆过滤、安全合计、动态阈值、检查机制和改进的Dice相似度函数的多方PPRL近似方法.首先利用布隆过滤将各数据源中的每条记录信息转换成由0和1组成的位数组；然后计算每个对应位置bit 1所占的比率，并利用动态阈值和检查机制来判定匹配成功的位置；最后通过改进的Dice相似度函数计算出记录间的相似度，进而判断记录间是否匹配成功.本文实验证明文中提出的方法具有较好的可扩展性，并且在保证查准率的同时，比已有的多方近似PPRL方法具有更高的容错性.

Multi-Party Privacy-Preserving Record Linkage Approach

Multi-Party Privacy-Preserving Record Linkage is the process of identifying records that correspond to the same real-world entities across several databases without revealing any sensitive information about these entities. With the increasing amount of data and the existing of real-world data quality issues (such as spelling errors, wrong order and so on), scalability and fault tolerance of PPRL have become the main challenges. At present, most of the existing multi-party PPRL methods are exact match without fault-tolerant. There are a few other PPRL approximate methods with fault-tolerant, but when dealing with the data existing quality issues, due to the low fault-tolerance and too much time cost they cannot effectively find out the common entities between databases. Therefore, we propose a multi-party PPRL approximate approach combined with bloom filter, secure summation, dynamic threshold, check mechanism and improved Dice similarity function. First, bloom filter is used to convert each record in the databases to an array of 1 and 0. Then we calculate ratio of bit 1 for each corresponding position, and use dynamic threshold and check mechanism to determine matched position. Finally we calculate the similarity between records by improved Dice similarity function, and judge whether records matched. In this paper, experimental results show the proposed method has good scalability and higher fault tolerance than the existing multi-party PPRL approximate method with good precision.