Personalized Differential Privacy Federated Learning Model for Internet of Vehicles
Long Zhang, Jinxin Zuo, Yueming Lu, Shihong Zou · 2024
To solve the issue of the differential privacy protection mechanism in federated learning (FL) for Internet of Vehicles (IoV), which only focuses on fixed privacy budget allocation methods and ignores the dynamic allocation of privacy budgets based on the risk of user data privacy leakage, this paper proposes a personalized differential privacy federated learning model based on privacy metrics. In each round of federated training, vehicle nodes simulate membership inference attacks to measure the performance of the attack model using confidence metrics. The attack results are then transformed into privacy risk scores, quantifying the degree of privacy leakage of the training nodes. Subsequently, the vehicle nodes establish a multilevel privacy budget allocation method, allowing users to adjust the magnitude of noise added to the model parameters based on the privacy risk level before uploading. This enables users to achieve dynamic protection during the federated learning training phase. Experimental results show that the personalized differential privacy algorithm based on privacy metrics improves the accuracy of the model compared to the method of uniformly setting the privacy budget.