《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (12): 3790-3798.DOI: 10.11772/j.issn.1001-9081.2022121831

• 网络空间安全 • 上一篇    下一篇

基于生成对抗网络的联邦学习中投毒攻击检测方案

陈谦, 柴政, 王子龙(), 陈嘉伟   

  1. 西安电子科技大学 网络与信息安全学院,西安 710071
  • 收稿日期:2022-12-13 修回日期:2023-05-09 接受日期:2023-05-10 发布日期:2023-05-25 出版日期:2023-12-10
  • 通讯作者: 王子龙
  • 作者简介:陈谦(1993—),男,陕西西安人,博士研究生,主要研究方向:隐私保护、机器学习、联邦学习
    柴政(1999—),男,黑龙江大庆人,硕士研究生,主要研究方向:联邦学习、投毒攻击
    陈嘉伟(1998—),男,陕西西安人,硕士研究生,主要研究方向:联邦学习、强化学习。
  • 基金资助:
    国家自然科学基金资助项目(62172319)

Poisoning attack detection scheme based on generative adversarial network for federated learning

Qian CHEN, Zheng CHAI, Zilong WANG(), Jiawei CHEN   

  1. School of Cyber Engineering,Xidian University,Xi’an Shaanxi 710071,China
  • Received:2022-12-13 Revised:2023-05-09 Accepted:2023-05-10 Online:2023-05-25 Published:2023-12-10
  • Contact: Zilong WANG
  • About author:CHEN Qian, born in 1993, Ph. D. candidate. His research interests include privacy preservation, machine learning, federated learning.
    CHAI Zheng, born in 1999, M. S. candidate. His research interests include federated learning, poisoning attack.
    CHEN Jiawei, born in 1998, M. S. candidate. His research interests include federated learning, reinforcement learning.
  • Supported by:
    the National Natural Science Foundation of China(62172319)

摘要:

联邦学习(FL)是一种新兴的隐私保护机器学习(ML)范式,然而它的分布式的训练结构更易受到投毒攻击的威胁:攻击者通过向中央服务器上传投毒模型以污染全局模型,减缓全局模型收敛并降低全局模型精确度。针对上述问题,提出一种基于生成对抗网络(GAN)的投毒攻击检测方案。首先,将良性本地模型输入GAN产生检测样本;其次,使用生成的检测样本检测客户端上传的本地模型;最后,根据检测指标剔除投毒模型。同时,所提方案定义了F1值损失和精确度损失这两项检测指标检测投毒模型,将检测范围从单一类型的投毒攻击扩展至全部两种类型的投毒攻击;设计阈值判定方法处理误判问题,确保误判鲁棒性。实验结果表明,在MNIST和Fashion-MNIST数据集上,所提方案能够生成高质量检测样本,并有效检测与剔除投毒模型;与使用收集测试数据和使用生成测试数据但仅使用精确度作为检测指标的两种检测方案相比,所提方案的全局模型精确度提升了2.7~12.2个百分点。

关键词: 联邦学习, 投毒攻击, 生成对抗网络, F1值损失, 精确度损失, 阈值判定方法

Abstract:

Federated Learning (FL) emerges as a novel privacy-preserving Machine Learning (ML) paradigm. However, the distributed training structure of FL is more vulnerable to poisoning attack, where adversaries contaminate the global model through uploading poisoning models, resulting in the convergence deceleration and the prediction accuracy degradation of the global model. To solve the above problem, a poisoning attack detection scheme based on Generative Adversarial Network (GAN) was proposed. Firstly, the benign local models were fed into the GAN to output testing samples. Then, the testing samples were used to detect the local models uploaded by the clients. Finally, the poisoning models were eliminated according to the testing metrics. Meanwhile, two test metrics named F1 score loss and accuracy loss were defined to detect the poisoning models and extend the detection scope from one single type of poisoning attacks to all types of poisoning attacks. Besides, a threshold determination method was designed to deal with misjudgment, so that the robust of misjudgment was confirmed. Experimental results on MNIST and Fashion-MNIST datasets show that the proposed scheme can generate high-quality testing samples, and then detect and eliminate poisoning models. Compared with the global models trained with the detection scheme based on directly gathering test data from clients and the detection scheme based on generating test data and using test accuracy as the test metric, the global model trained with the proposed scheme has significant accuracy improvement from 2.7 to 12.2 percentage points.

Key words: Federated Learning (FL), poisoning attack, Generative Adversarial Network (GAN), F1 score loss, accuracy loss, threshold determination method

中图分类号: