基于车雾环境的联邦学习动态隐私保护方案

Dynamic Privacy Protection Scheme of Federated Learning in Vehicular Fog Environment

当前已经有研究将雾环境与联邦学习结合应用在车联网隐私保护中,但是缺乏对车辆移动性可能导致隐私需求改变的问题的考虑。为此,文中基于区域内车辆终端数目,提出了在不同的隐私需求下实施不同的隐私保护和效率调整的方案,在同态加密方案中进行双重加密聚合并且动态调整本地迭代次数,在差分隐私方案中动态调整每轮云聚合与雾聚合次数。实验表明,在区域内车辆终端数不同的情况下,本方案满足在隐私计算的同时保持较高精度。

At present,some studies combine federated learning with fog environment in the privacy protection of Internet of vehicles,but lack of consideration of the problem that the mobility of vehicle may lead to the change of privacy requirements.For this,we propose implementing different privacy protection and efficiency adjustment schemes under different privacy requirements,based on the number of vehicle terminals in the area.In the homomorphic encryption scheme,we perform double encryption aggregation and adjust the number of local iterations in this paper.In the differential privacy scheme,the number of cloud aggregation and fog aggregation in each round is dynamically adjusted.Finally,the experience shows that even if the number of vehicle terminals in the area is different,the scheme can maintain high accuracy while achieving privacy calculation.