Differentially private federated learning: A systematic review (ACM Survey); Adap dp-fl: Differentially private federated learning with adaptive noise (TrustCom
Differentially private federated learning: A systematic review (ACM Survey); Adap dp-fl: Differentially private federated learning with adaptive noise (TrustCom
Person Website(个人主页): https://Jefffffffu.github.io/
bilibili(论文视频分享):https://space.bilibili.com/80356866/video
关于DP-FL更完整更系统的总结,请查阅:https://arxiv.org/abs/2405.08299
推荐一个DP社区:https://differentialprivacy.org/
以上的分类总结和代码来自于以下论文,如果你从以上总结得到了帮助或者你有使用相关代码,请您在文章中引用以下论文。这对我非常重要,谢谢。
@inproceedings{fu2022adap,
title={Adap dp-fl: Differentially private federated learning with adaptive noise},
author={Fu, Jie and Chen, Zhili and Han, Xiao},
booktitle={2022 IEEE International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom)},
pages={656--663},
year={2022},
organization={IEEE}
}@article{fu2024differentially,
title={Differentially private federated learning: A systematic review},
author={Fu, Jie and Hong, Yuan and Ling, Xinpeng and Wang, Leixia and Ran, Xun and Sun, Zhiyu and Wang, Wendy Hui and Chen, Zhili and Cao, Yang},
journal={arXiv preprint arXiv:2405.08299},
year={2024}
}Table of Contents
CDP(central DP)有一个完全可信的中心方,敌手是外界。而LDP(local DP)认为中心方是诚实但好奇的。
| Title | Team/Main Author | Venue and Year | Key Description |
|---|---|---|---|
| Differential privacy | Cynthia Dwork | ICALP/2006 | 首次提出差分隐私的定义 |
| Programming Differential Privacy (Book) | Joseph P. Near and Chiké Abuah | 2021 | 讲诉了DP的概念定理和机制等,并附有简单代码呈现(简单入门推荐)【Link】 |
| The Algorithmic Foundations of Differential Privacy(Book) | Cynthia Dwork | 2014 | DP的定义理论,高级组合和相关机制等的完整证明推导(更加理论)【拉普拉斯、严格差分、高斯机制、松弛差分】 |
| Differential Privacy From Theory to Practice (Book) | Ninghui Li | 2014 | 除了一些基本定理和机制,用了具体的实际例子讲诉了DP的用法及DP的伦理探讨(更加实用化)【Chapter1、Chapter2】,原作者讲解 |
| Title | Team/Main Author | Venue and Year | Key Description |
|---|---|---|---|
| 本地化差分隐私研究综述 | 中国人民大学 | Journal of Software/2018 | 介绍了本地化差分隐私的原理与特性,总结和归纳了LDP当前研究工作,重点阐述了频数统计、均值统计的LDP机制设计【vedio】 |
| RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response | ACM SIGSAC /2014 | 1.RAPPOR的分类和对应扰动方法。2.Differential Privacy loss of RAPPOR(主要是多维) | |
| Locally Differentially Private Protocols for Frequency Estimation | Purdue University | USENIX/2017 | 1.提出了一个Pure LDP Protocols,并基于Pure LDP Protocols给出了其方差和频数估计的公式。2.提出OUE(UE=basic RAPPOR),给出了q的最佳选择和方差。【vedio】 |
| Collecting High-Dimensional and Correlation-Constrained Data with Local Differential Privacy | Rong Du | sensor, mesh and ad hoc communications and networks/2021 | 该文章重要的是针对LDP的高维度情况,本篇重点在于推断多维RR的概率关系: 1、进行属性之间相关性的度量(首先会属性会分为不同的簇,簇之间独立不相干,簇之内的属性相关),定义了p这个变量定义来定理相关性(基于协方差) 2、提出UDLDP,基于以上在单个属性上定义LDP,并提出CBP |
| Collecting and Analyzing Data from Smart Device Users with Local Differential Privacy | Thông T. Nguyên et al. | arxiv/2016 | 提出了Harmony,用于包含数值和类别属性的多维数据的LDP下的均值和频数统计。主要是连续型数据直接随机的对称扰动成两个相反数,然后保证均值无偏,误差边界比用Lap小。 |
| Collecting and Analyzing Multidimensional Data with Local Differential Privacy | Ning Wang et al. | ICDE/2019 | 核心是单维数据LDP的收集,多维是一个简单扩展。单维下文章发现了拉普拉斯加噪和DM(Duchi)的优缺点,两个方法随着eps的增大有一个交点。文章结合两个方法的优点提出PM方法,随后引入alpha参数升级为HM得到更小的误差边界 【vedio】 |
| A Comprehensive Survey on Local Differential Privacy | Wuhan University | Security and Communication Networks/2020 | 目前看起来较为新较为全面的LDP综述,对于LDP的各种机制和应用都有概括。可以反复进行参考 |
| Utility-Optimized Local Differential Privacy Mechanisms for Distribution Estimation | Takao Murakami | USENIX Security/2019 | 针对LDP下的不同回答(属性),文章认为有些是敏感的需要保护,有些是不敏感的不需要保护,由此提出一种新的LDP叫ULDP,ULDP使得在同样的预算下做更少的扰动,使得方差误差更小。【vedio】 |
| Title | Team/Main Author | Venue and Year | Key Description |
|---|---|---|---|
| Differential privacy | Cynthia Dwork | ICALP/2006 | 首次提出差分隐私的定义 |
| Programming Differential Privacy (Book) | Joseph P. Near and Chiké Abuah | 2021 | 讲诉了DP的概念定理和机制等,并附有简单代码呈现(简单入门推荐) |
| The Algorithmic Foundations of Differential Privacy(Book) | Cynthia Dwork | 2014 | DP的定义理论,高级组合和相关机制等的完整证明推导(更加理论)【拉普拉斯、严格差分、高斯机制、松弛差分】 |
| Differential Privacy From Theory to Practice (Book) | Ninghui Li | 2014 | 除了一些基本定理和机制,用了具体的实际例子讲诉了DP的用法及DP的伦理探讨(更加实用化)【Chapter1、Chapter2】,原作者讲解 |
| The Bounded Laplace Mechanism in Differential Privacy | IBM | 2018 | 提出了阈值限定的重加噪拉普拉斯机制无法满足纯DP,但是可以满足松弛DP,文章给出了噪声系数需要满足一个式子(该式子和区间,eps,delta相关)即可满足(eps,delta)-DP。这边的松弛DP不是用的RDP,而是直接在后面加一个delta来放缩sigma。 |
此前的MA(monments accountant)目前来看就是RDP,区别在于RDP从一开始的散度就用了高阶矩进行度量,而MA是在分布上进行高阶矩度量,不过最后的表达形式几乎一致,两者思想也一致。
| Title | Team/Main Author | Venue and Year | Key Description |
|---|---|---|---|
| Deep Learning with Differential Privacy | Martín Abadi | CCS/2016 | 首次提出高斯下采样机制并利用Moments accountant技术进行隐私损失度量 【vedio】 |
| Rényi Differential Privacy | Ilya Mironov | Security Foundations Symposium/2017 | Rényi 差分隐私的定义、性质和结论。主要关注里面的高斯机制(无下采样),组合性质及转为DP的公式。【vedio】 |
| Subsampled Rényi Differential Privacy and Analytical Moments Accountant | Yu-Xiang Wang | AISTATS /2018 | RDP和MA的直观对比,主要给出了一个RDP高斯下采样的一个隐私损失上界,有很多论文引用这个上界的结论。 |
| Rényi Differential Privacy of the Sampled Gaussian Mechanism | Ilya Mironov | 2019 | 给出了Rényi Differential Privacy of the Sampled Gaussian Mechanism 各种情况的隐私损失公式,里面的3.3的公式目前应用于各个开源库的隐私计算中(更为简洁) |
| Hypothesis Testing Interpretations and Rényi Differential Privacy | Balle B | AISTATS/2020 | 定理21给出了更为紧凑的RDP转DP的公式(目前开源库opacus应用的是这个转换) |
| Title | Team/Main Author | Venue and Year | Key Description |
|---|---|---|---|
| Concentrated Di |
No open issues yet, or sync has not completed.