论文总字数：33017字

摘 要

新时代以来，移动设备已成为了一种常态，纷至沓来的各种移动设备应用令人应接不暇。由于定位技术的不断发展，基于个人位置信息的各种服务（LBS）也成为了人们手机中必不可少的一部分，为人们的出行提供了便捷和安全保障，如规划当前地点与目的地之间的路径，搜索周边的服务或者担任自己迷路时的救命稻草。然而，在LBS为人们带来无限便利的同时，隐私的泄露也会随之而来。无时不刻的使用LBS会让自己的行踪完全暴露在别人的手中，为他人所利用，这已经对个人的隐私造成了很大的威胁。因此，设计一个合适的隐私保护机制是迫在眉睫的，这能够很好地在保证用户的服务体验的基础上保护用户的位置隐私安全。

本文主要研究设计位置隐私保护机制，保护用户发出LBS请求时的位置隐私。由生成式对抗网络（GAN）得到启发，具体工作如下：

首先，为了设计出能够真正符合真实用户需求的机制，首先需要收集并产生合适的训练集。从Gowalla数据集中获取到全世界用户提交的地理位置信息数据，选择某个地区的所有用户数据作为原始数据集，并添加100次的拉普拉斯噪声用于混淆其真实位置。为了便于后续的研究，将所有的位置数据进行规格化处理。

其次，受到GAN生成式对抗网络的启发，设计出由产生器G和分类器C组成的对抗网络，其损失函数即代表着对隐私性与效用性的评估。对隐私的量化采用的是相互信息（MI）的新方法，摒弃了传统的交叉熵（CE），使最终的收敛结果能够最大化的折衷隐私性和实用性。该网络亦能根据不同的用户需求调整其对于隐私性和实用性之间的权重关系。

最后，使用梯度下降的方法训练网络并调试网络内的各种参数，获取训练过程中的部分生成器和最优的生成器G，通过作图观察其逐步的变化过程。将分类器C与理想最佳Bayes分类器（最强大的对手）进行比较后对机制的表现效果进行评估。

通过本文设计的基于生成式对抗网络的对个人位置隐私的保护机制可以有效地达成对用户隐私的保护同时保证用户获取相应的LBS服务，并且该机制能够简单的部署在移动端进行离线服务。

关键词：生成式对抗网络 LBS 位置隐私 相互信息 拉普拉斯噪声

Design of Location Privacy Protection Mechanism based on Generative Adversarial Network

Abstract

Since the new era, mobile devices have become the norm, and the applications of various mobile devices are overwhelming. Due to the continuous development of positioning technology, personal location based information service (LBS) has become an indispensable part in the mobile phone, providing convenience and safety for people, such as planning routes between the current location and the destination path, searching for the service around or serving as a lifeline when he or she gets lost. However, while LBS brings infinite convenience to people, the disclosure of privacy will also follow. The constant use of LBS will make their whereabouts completely exposed to others and be utilized by others, which has posed a great threat to personal privacy. Therefore, it is urgent to design an appropriate privacy protection mechanism, which can protect users' location privacy and security on the basis of ensuring users' service experience.

This paper mainly studies the design of location privacy protection mechanism to protect users' location privacy when making LBS requests. Inspired by the generative adversarial network (GAN), the specific work is as follows:

First, in order to design a mechanism that truly meets the needs of a real user, you need to collect and generate the right training set. The geographic location information data submitted by users around the world is obtained from the Gowalla data set, all user data in a certain region is selected as the original data set, and Laplace noise is added 100 repetitions to confuse its real location. In order to facilitate the follow-up study, all location data are scaled.

Secondly, inspired by GAN generative adversarial network, an adversarial network composed of generator G and classifier C is designed, whose loss function represents the evaluation of privacy and utility. The method to quantify privacy adopts a new method of mutual information (MI), rejecting the traditional cross entropy (CE), so that the final convergence result can maximize the compromise of privacy and practicability. The network can also adjust its weight between privacy and utility according to different user needs.

Finally, using the method of gradient descent to train and adjust the various parameters within the network, and then get some of the generators through the training process as well as the optimal generator G in the end, its gradual change is able to be observed by plots. Compare the classifier C with the ideal best Bayes classifier (the most powerful opponent) and then evaluate the performance of the mechanism.

The mechanism of personal location privacy protection based on generative adversarial network designed in this paper can effectively protect users' privacy and guarantee users to obtain corresponding LBS services. Moreover, this mechanism can be easily deployed on mobile terminals for offline services.

Keywords: generative adversarial network; LBS; location privacy; mutual information; Laplace noise

目录

摘 要 I

Abstract III

目录 V

第一章 绪论 1

1.1 课题背景 1

1.2 课题意义 2

1.3 研究动机 3

1.4 本文主要工作 4

1.5 论文结构 4

第二章 预备知识 6

2.1 位置隐私保护 6

2.1.1 系统架构 7

2.1.2 度量指标 8

2.1.3 保护方法 8

2.2 生成式对抗网络 10

2.2.1 基本概念 10

2.2.2 实现原理 10

2.2.3 典型应用 11

2.3 本章小结 11

第三章 系统模型 12

3.1 假设与变量定义 12

3.2 系统基本设定 13

3.2.1 量化效用方式 14

3.2.2 使用相互信息量化隐私 14

3.2.3 博弈的设定以及均衡策略 15

3.2.4 采用Bayes误差评估机制 16

3.3 原始数据获取与处理 16

3.3.1 原始数据集的获取 17

3.3.2 原始数据的规格化 17

3.4 本章小结 18

第四章 神经网络实现 19

4.1 神经网络框架 19

4.1.1 符号说明 19

4.1.2 基本原理 19

4.2 网络实现中细节 21

4.2.1 基本模型 21

4.2.2 相互信息 23

4.2.3 训练周期与批次大小 26

4.2.4 其他细节 26

4.3 本章小结 27

第五章 在数据集上的实验 29

5.1 平面拉普拉斯机制 29

5.2 Bayes误差估计 30

5.3 Gowalla数据集上的实验 30

5.3.1 宽松的效用约束 31

5.3.2 更严格的效用约束 33

5.4 本章小结 35

第六章 总结与展望 36

参考文献 37

致谢 41

绪论

1.1 课题背景

数据分析对现代公司至关重要，因此，人们对收集和处理各种各样的个人信息有着极大的兴趣。另一方面，个人往往愿意提供他们的数据，以换取更好的服务和经验。然而，这样的个人信息泄露可能会被用来对付他们:它可能会影响一切，从人际关系到找到工作，或获得贷款资格，甚至更糟。

请支付后下载全文，论文总字数：33017字