| 摘要: |
| 针对车辆自组织网络（Vehicular Ad-hoc Networks,VANETs）中联邦学习（Federated Learning,FL）的隐私泄露与身份认证缺失问题，提出了一种面向VANETs身份持续认证的隐私保护联邦学习（Privacy-preserving Federated Learning for Continuous Authentication，PPFLCA）方案，通过车辆-路侧单元（Road Side Units，RSUs）-聚合中心三层结构来更新全局模型。具体地，PPFLCA首先进行车辆和RSUs的注册，并为车辆生成假名；其次，利用密钥协商协议生成共享密钥，结合伪随机生成器（Pseudorandom Generator，PRG）生成随机掩码，并采用16-bit定点数优化加密本地梯度；最后，每次迭代对车辆和RSUs身份和本地梯度进行持续认证。安全分析表明，PPFLCA可以满足隐私保护需求，同时可以抵抗基本攻击和重放攻击。实验分析表明，PPFLCA在MNIST和CIFAR-10数据集上分类准确率分别达94.8%和44.4%（与明文训练误差小于0.5%），在梯度参数为5×103时，车辆梯度加密耗时18 ms，车辆通信开销仅0.367 KB。 |
| 关键词: 车辆自组织网络(VANETs) 联邦学习 隐私保护 身份认证 |
| DOI：10.20079/j.issn.1001-893x.250113002 |
|
| 基金项目:国家自然科学基金资助项目（62072133）；温州市重大科技创新攻关项目（ZG2023028） |
|
| Privacy-preserving Federated Learning for Continuous Authentication Scheme in VANETs |
| XIA Yuanjun,LI Lihui,REN Shuyang,YU Lele,LIU Yining |
| (1.School of Data Science and Artificial Intelligence,Wenzhou University of Technology,Wenzhou 325027,China;2.School of Computer and Information Security,Guilin University of Electronic Technology,Guilin 541004,China) |
| Abstract: |
| For the issues of privacy leakage and identity authentication deficiency in federated learning(FL) for vehicular ad-hoc networks(VANETs),a privacy-preserving federated learning for continuous authentication(PPFLCA) scheme in VANETs is proposed.The scheme updates global models via a three-tiered architecture comprising vehicles,road side units(RSUs),and an aggregation center.Specifically,PPFLCA initiates vehicle and RSU registration while generating pseudonyms for vehicles.Subsequently,a key agreement protocol is employed to generate shared keys,which is combined with a pseudorandom generator(PRG) to produce random masks,and the scheme further adopts 16-bit fixed-point numbers to encrypt and optimize local gradients.Finally,PPFLCA continuously authenticates the identities of vehicles and RSUs,as well as their local gradients in each iteration.Security analysis demonstrates that PPFLCA effectively fulfills privacy preservation requirements while resisting fundamental attacks and replay attacks.Experimental evaluations reveal classification accuracies of 94.8% and 44.4% on MNIST and CIFAR-10 datasets respectively(with error margin less than 0.5% compared with plaintext training).With gradient parameters of 5×103,the scheme achieves 18 ms encryption latency per vehicle and maintains a low communication overhead of 0.367 KB per vehicle. |
| Key words: vehicular ad-hoc networks(VANETs) federated learning privacy preserving identity authentication |
《电讯技术》编辑部 