Decoupling Statistical Trends from Data Volume on LDP-Based Spatio-Temporal Data Collection
Decoupling Statistical Trends from Data Volume on LDP-Based Spatio-Temporal Data Collection
复制标题
DOI：
10.1109/fnwf55208.2022.00053
复制
发表时间：
2022-10
期刊：
影响因子：
--
通讯作者：
Taisho Sasada;Yuzo Taenaka;Y. Kadobayashi
中科院分区：
文献类型：
--
作者：
Taisho Sasada;Yuzo Taenaka;Y. Kadobayashi
时空数据对于城市规划、流行病学和自然灾害等各种应用都很有用，但由于涉及到人们的隐私，会导致家庭/工作地址等私人信息的暴露。基于局部差分隐私（LDP）的处理是一种很有前途的技术，用于删除敏感信息的时空数据。基于LDP的处理添加了一定量的噪声，以使每条数据无法区分，同时保持其内在价值。然而，LDP容易受到数据放大的影响。当数据存储从任何设备接收数据时，数据存储仅将接收到的数据追加到现有数据。这允许任何人向数据中注入任何数量的数据，并操纵整个数据的趋势。为了解决这个问题，我们设计了一种数据收集方法，使数据存储能够收集来自每个设备的数据的统计趋势，而不管数据量如何。我们利用一个不经意传输（OT）协议，在接收端，数据存储执行数据包采样。这种采样使得能够收集统计趋势，但需要调整LDP处理，因为噪声量是由数据存储接收每一条LDP处理的数据的假设确定的。然后，我们提出了一种调整方法的基础上的欧几里德算法的LDP为基础的过程。我们进行了定性和实验开销分析，并表明，该方法的统计趋势和数据量之间的关系。我们还表明，处理负载在智能手机和物联网等小型设备上是可以接受的。
Spatio-temporal data is useful for various applications such as urban planning, epidemiology, and natural disasters, but causes exposure of private information, such as home/workplace addresses, because it involves people's trajec-tories. Local Differential Privacy (LDP) based processing is a promising technology for removing sensitive information from spatio-temporal data. A LDP-based processing adds a certain amount of noise to make each piece of data indistinguishable while keeping its intrinsic value. However, LDP is vulnerable to data amplification. When a data store receives data from any device, the data store only appends the received data to existing data. This allows anyone to inject any amount of data into the data and manipulate the trend of the whole data. To tackle this problem, we design a data collection method enabling a data store to collect statistical trends of data from every device irrespective of the data volume. We utilize an Oblivious Transfer (OT) protocol that performs a packet sampling at the reception side, the data store. This sampling enables the collection of statistical trends but requires adjusting LDP processing because the amount of noise is determined by the assumption that the data store receives every piece of LDP-processed data. We then propose an adjustment method for LDP-based process based on the Euclidean algorithm. We conducted qualitative and experimental overhead analysis and showed that the proposed method decouples the relationship between statistical trend and data volume. We also show the processing load can be acceptable on small devices such as smartphones and loT.