The large-scale model (LSM) can handle large-scale data and complex problems, effectively improving the intelligence level of urban intersections. However, the traffic conditions at intersections are becoming increasingly complex, so the intelligent intersection LSMs (I²LSMs) also need to be continuously learned and updated. The traditional cloud-based training method incurs a significant amount of computational and storage overhead, and there is a risk of data leakage. The combination of edge artificial intelligence and federated learning provides an efficient and highly privacy protected computing mode. Therefore, we propose a hierarchical hybrid distributed training mechanism for I²LSM. Firstly, relying on the intelligent intersection system for cloud-network-terminal integration, we constructed an I²LSM hierarchical hybrid distributed training architecture. Then, we propose a hierarchical hybrid federated learning (H²Fed) algorithm that combines the advantages of centralized federated learning and decentralized federated learning. Further, we propose an adaptive compressed sensing algorithm to reduce the communication overhead. Finally, we analyze the convergence of the H²Fed algorithm. Experimental results show that the H²Fed algorithm reduces the communication overhead by 21.6% while ensuring the accuracy of the model.

In crowdsourced federated learning, differential privacy is commonly used to prevent the aggregation server from recovering training data from the models uploaded by clients to achieve privacy preservation. However, improper privacy budget settings and perturbation methods will severely impact model performance. In order to achieve a harmonious equilibrium between privacy preservation and model performance, we propose a novel architecture for crowdsourced federated learning with personalized privacy preservation. In our architecture, to avoid the issue of poor model performance due to excessive privacy preservation requirements, we establish a two-stage dynamic game between the task requestor and clients to formulate the optimal privacy preservation strategy, allowing each client to independently control privacy preservation level. Additionally, we design a differential privacy perturbation mechanism based on weight priorities. It divides the weights based on their relevance with local data, applying different levels of perturbation to different types of weights. Finally, we conduct experiments on the proposed perturbation mechanism, and the experimental results indicate that our approach can achieve better global model performance with the same privacy budget.