A Novel Federated Learning Model Aggregation Method Empowered by Coordination of Computing and Networking

• Published in: 2023 International Conference on Networks, Communications and Intelligent Computing (NCIC) pp. 61 - 67
• Main Authors: Wang, Shoufeng, Chen, Hua-Min, Guo, Jianchao, Ouyang, Ye, Li, Fan, Chen, Xuan, Li, Yun, Ma, Leiming, Sun, Jie, Wang, Shuling
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.11.2023
• Subjects: 6G mobile communication; Cloud computing; Computational modeling; Coordination of Computing and Networking; Data privacy; Federated learning; Industries; Training
• DOI: 10.1109/NCIC61838.2023.00016

Federated learning is a machine learning training paradigm that enables model training under data privacy constraints. Its distinguishing capability is the ability to train models without requiring participating parties to share private data. However, model training and aggregation in federated learning is communication and computation intensive. Relying solely on cloud computing is insufficient to balance the trade-offs in model aggregation and training given the distributed nature of federated learning. Coordination of Computing and Networking (CNC), an emerging infrastructure, incorporates massive cloud, edge / end computing into the communication network enabled by ubiquitous 6G connectivity. This forms an integrated infrastructure encompassing both networking and ubiquitous computing. CNC can improve performance for computing-intensive workloads, time-sensitive applications and distributed computing tasks via on-demand computing invocation. This paper studies how to exploit the ubiquitous distributed computing resources of computational networks to facilitate federated learning model training and aggregation. Challenges addressed include communication overhead from repeated model aggregations across participating nodes and the central node, computing overhead for local model training at participating nodes and global model aggregation at the central node, and improving computational network resource utilization efficiency throughout this process.