文献类型：期刊文献

中文题名：抗拜占庭攻击的梯度净化联邦自适应学习算法

英文题名：Gradient purification federated adaptive learning algorithm for Byzantine attack resistance

作者：杨辉[1];邱子游[1];李中美[2];朱建勇[1]

机构：[1]华东交通大学电气与自动化工程学院,江西南昌330013;[2]华东理工大学信息科学与工程学院,上海200237

年份：2024

卷号：45

期号：S1

起止页码：1

中文期刊名：通信学报

外文期刊名：Journal on Communications

收录：;北大核心:【北大核心2023】；CSCD:【CSCD2023_2024】；

基金：国家自然科学基金资助项目(No.62363010,No.61733005);工业控制技术国家重点实验室开放课题基金资助项目(No.ICT2024B50)。

语种：中文

中文关键词：联邦学习;拜占庭攻击;投毒攻击;模型鲁棒性;工业大数据

外文关键词：federated learning;Byzantine attack;poisoning attack;model robustness;industrial big data

摘要：在工业大数据之下,数据安全和隐私保护是关键挑战之一。传统的数据共享和模型训练方法在应对数据泄露和恶意攻击(尤其是复杂的拜占庭攻击和投毒攻击)时效果有限,因为传统联邦学习通常假定所有参与方都是可信的,这使得模型在遭遇投毒攻击时性能显著下降。为解决这个问题,本文提出一种抗拜占庭攻击的梯度净化联邦自适应学习算法,通过滑动窗口梯度过滤器和符号聚类过滤器识别恶意梯度,滑动窗口方法检测异常梯度,而符号聚类则根据梯度方向一致性筛选出偏离的对抗性梯度,经过过滤后,使用基于权重的自适应聚合规则对剩余的可信梯度进行加权聚合,动态调整参与方梯度的权重,降低恶意梯度的影响,从而增强模型的鲁棒性。实验结果显示,尽管新型投毒攻击的强度更高,但所提算法能有效防御这些攻击且减轻模型性能的损失。相比于传统防御算法,所提算法不仅提高了模型的准确性,还提升了其安全性。
In the context of industrial big data,data security and privacy are key challenges.Traditional data-sharing and model-training methods struggle against risks like Byzantine and poisoning attacks,as federated learning typically assumes all participants are trustworthy,leading to performance drops under attacks.To address this,a Byzantine-resilient gradient purification federated adaptive learning algorithm was proposed.The malicious gradients were identified through a sliding window gradient filter and a sign-based clustering filter.The sliding window method detected anomalous gradients,while the sign-based clustering filter selected adversarial gradients based on the consistency of gradient directions.After filtering,a weight-based adaptive aggregation rule was applied to perform weighted aggregation on the remaining trustworthy gradients,dynamically adjusting the weights of participant gradients to reduce the impact of malicious gradients,thereby enhancing the model’srobustness.Experimental results show that despite the increased intensity of new poisoning attacks,the proposed algorithm effectively defends against these attacks while minimizing the loss in model performance.Compared to traditional defense algorithms,it not only improves model accuracy but also enhances its security.