基于联邦学习的隐私保护区域交通流量预测模型

Privacy-Preserving Regional Traffic Flow Prediction Model Based on Federated Learning

区域交通流量预测是智慧交通系统的一项重要功能。联邦学习可以支持多位置服务提供商(Location Service Provider,LSP)的联合训练,使得训练数据集可以更加全面地覆盖整个区域的交通流量,提高预测准确率。但是,当前基于联邦学习的区域交通流量预测方案存在车辆数据去重、训练节点背叛以及隐私泄露等问题。为此,构建了基于联邦学习的隐私保护区域交通流量预测(Privacy-Preserving Regional Traffic Flow Prediction based on Federated Learning,PPRTFP-FL)模型。模型采用中心部署架构,由联邦中央服务器协调各个LSP联合完成模型的训练,并对全局模型进行梯度聚合与模型更新;采用交叉评价加权聚合的策略来防御部分不可信节点对全局模型的恶意攻击,提升了全局模型的鲁棒性;预测阶段使用同态加密聚合算法,各LSP在不泄露自身运营数据的情况下实现了更准确的流量预测。利用相关数据集进行测试,测试结果表明当训练数据集覆盖区域流量充分的情况下,本模型相比本地模型的预测准确率有明显的提升。对模型进行不同比例的恶意节点攻击实验,由实验结果可知,系统在存在恶意节点情况(当恶意节点数量小于总节点数量50%时)下仍具备较好的防御效果。

Regional traffic flow prediction is an important function of intelligent transportation systems.Federated learning can support joint training of multiple Location Service Providers(LSPs),making the training dataset more comprehensive in covering the traffic flow of the entire region and improving prediction accuracy.However,the current regional traffic flow prediction scheme based on federated learning has issues such as vehicle data deduplication,training node betrayal,and privacy leakage.A Privacy-Preserving Regional Traffic Flow Prediction model based on Federated Learning(PPRTFP-FL)is proposed.The central deployment architecture was adopted,and the federal central server coordinated the LSPs to jointly complete the model training,and carried out gradient aggregation and model update for the global model.The cross-evaluation weighted aggregation strategy is used to defend against malicious attacks from partially untrusted nodes,which improves the robustness of the global model.In the prediction phase,the homomorphic encryption aggregation algorithm is used to achieve more accurate traffic prediction for LSPs without disclosing their own operating data.The test results show that the prediction accuracy of this model is significantly improved compared with the local model when the training data set covers sufficient area traffic.The experimental results show that the system has a good defense effect when the number of malicious nodes is less than 50%of the total number of nodes.