Joint Contract Design and Task Reorganization for Semi-Decentralized Federated Edge Learning in Vehicular Networks

Federated edge learning (FEEL) emerges as a privacy-preserving paradigm to effectively integrate edge computing for the implementation of deep learning-based vehicular applications. Nevertheless, the incentive mechanism for vehicles participating in varied learning tasks, has not been well explored yet. In this paper, software-defined network (SDN) technology is adopted for the training control among vehicles, and a novel FEEL framework, namely SDN-assisted semi-decentralized FEEL (SSD-FEEL) is investigated, where multiple edge servers collectively coordinate a large number of vehicular models from different learning tasks. By exploiting the low-cost and similar learning tasks among vehicles and edge servers, SSD-FEEL incorporates more training samples, while enjoying the flexibility of edge server assisted model aggregation. Aiming at motivating vehicles to actively participate in training while improving the model accuracy of multiple learning tasks, a joint contract design and task reorganization problem, combined with the evaluation of model convergence and contract performance, is formulated. Then, we propose a two-stage optimization algorithm incorporating iterative reward allocation and task matching, where the model parameters in different tasks are reconstructed according to the matching results with the mobility constraints. Extensive experiments conducted on multiple data sets validate that the proposed algorithm can achieve higher cluster utility and outperform the conventional multi-task FEEL schemes in terms of learning performance.