一种基于隐私保护下的多方记录链接方法

发布时间：2018-03-24 14:44

本文选题：记录链接　切入点：隐私保护　出处：《软件学报》2017年09期

【摘要】：多方隐私保护下的记录链接(privacy-preserving record linkage,简称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外,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,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.实验结果表明:所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.
[Abstract]:Privacy-preserving record link (PPRL) under multi-party privacy protection is a process in which the same entity in the real world is found from multiple data sources under privacy protection. Other information has not been leaked. With the increasing volume of data and the existence of real-world data quality problems (such as spelling errors, reverse order, etc.), the extensibility and fault-tolerance of multi-party PPRL methods are facing challenges. Most of the existing multiparty PPRL methods are accurate matching methods, but they are not fault-tolerant. There are also a few multiparty PPRL approximation methods that are fault-tolerant, but when dealing with data with quality problems, the fault tolerance is poor and the time cost is too high. Therefore, a multi-party PPRL approximation method combining Bron filtering, security total, dynamic threshold, checking mechanism and improved Dice similarity function is proposed. Each record information in each data source is converted into a bit array of 0 and 1 using the Bloom filter. Then, the ratio of bit 1 to each corresponding location is calculated. The dynamic threshold and checking mechanism are used to determine the matching position. Finally, the similarity between the records is calculated by the improved Dice similarity function. The experimental results show that the proposed method is more extensible and has higher fault tolerance than the existing multi-party approximate PPRL method.
【作者单位】：东北大学计算机科学与工程学院;
【基金】：国家自然科学基金(61472070,61672142) 国家重点基础研究发展计划(973)(2012CB316201)~~
【分类号】：TP309

【相似文献】