Enhancing privacy protection for vehicle trajectory prediction using differential privacy and federated learning

Intelligent Connected Vehicles (ICVs) are one of the rapidly developing directions in the field of autonomous driving and play an important role. At present, federated learning has gradually been used to support various collaborative applications in ICV scenarios, such as analyzing vehicle driving information to achieve trajectory prediction, voice activation control, etc. But research shows that federated learning systems still face the challenge of privacy breaches. These threats may lead to sensitive private data breaches, such as location information, driving conditions, etc. To address this issue, defense mechanisms such as homomorphic encryption and differential privacy have been introduced. However, homomorphic encryption requires enormous computational resources and additional communication overhead. In this work, we propose a new method that combines federated learning with differential privacy in vehicle trajectory prediction scenarios. Through experiments, we have demonstrated that at a noise factor of 0.01 in 100 epochs, the accuracy of vehicle trajectory prediction is comparable to the baseline, and at $\mathbf{0. 0 3}$, it can reach $\mathbf{7 7. 5 5 \%}$.