Privacy and Robustness in Federated Learning: Attacks and Defenses

Privacy and Robustness in Federated Learning: Attacks and Defenses abstract。

解决的问题

传统的中心式的机器学习方法无法有效处理隐私问题，联邦学习作为一种替代方案近年来得到发展。但是隐私问题并没有得到根本的解决，因此本文针对联邦学习面对的攻击和对应的防御手段对近5年的研究做出综述。

Introduction

根据分布式情况对联邦学习分类

• Horizontally federated learning, HFL
  • HFL to business, H2B
  • HFL to consumers, H2C
• Vertically federated learning, VFL
• Federated transfer learning, FTL

根据架构情况对联邦学习分类

• FL with Homogeneous Architectures
• FL with Heterogeneous Architectures

联邦学习面临的威胁

面临的威胁主要来自两个方面：

• malicious server
• adversarial participant

面临的主要攻击是：

• data poisoning
• model poisoning

Threat Model

从几个侧面对攻击者进行归类:

• Insider vs Outsider
• Training phase vs Inference phase
• Privacy: semi-honest vs malicious
• Robustnetss: untargeted vs targeted

Privacy Attacks

针对隐私的攻击有：

• Inferring Class Representatives
• Inferring Membership
• Inferring Properties
• Inferring Training Inputs and Labels

Defences against privacy attacks

• Homomorphic Encryption
• Secure Multiparty Computation, SMC
• Differential Privacy

Poisoning attacks

投毒攻击有：

• Untargeted, 无针对目标，旨在损害模型准确率和性能
• Targeted，有目标的，如后门攻击

Defences against poisoning attacks

Discussion and Promising Direction

• Curse of Dimensionality
• Rethinking Current Privacy Attacks
• Rethinking Current Defences
• Optimizing Defence Mechanism Deployment
• Test-phase Privacy in FL
• Test-phase Roubustness in FL
• Relationship with GDPR
• Threats and Protection of VFL and FTL
• Vulnerabilities to Free-riding Participants
• More Possibilities in FL with Heterogeneous Architectures
• Decentralized FL
• Efficient FL with Single Round Communication