Optimization of Sustainable Decentralized Federated Learning

While decentralized federated learning (DFL) has emerged as a promising paradigm for privacy-preserving and resilient collaborative training, its environmental footprint remains underexplored. Existing sustainability frameworks in federated learning mainly focus on device-level measurements such as CPU/GPU energy consumption or hardware efficiency. However, these approaches overlook the process-level dimensions of the DFL lifecycle, including model synchronization, aggregation, local training phases, communication overhead, and topology dynamics.

This thesis aims to design and implement a fine-grained sustainability quantification framework for DFL systems, extending beyond device-level metrics to capture process-level indicators such as computation, communication, aggregation, and synchronization costs. Based on this framework, the student will investigate optimization strategies that balance model performance with sustainability objectives, e.g., adaptive node participation, selective synchronization, or process-aware scheduling.

reference:

• Chao Feng, Alberto Huertas Celdran, Pedro Miguel Sanchez Sanchez, Lynn Zumtaugwald, Gerome Bovet, Burkhard Stiller: Assessing the Sustainability and Trustworthiness of Federated Learning Models; 21st International Conference on Network and Service Management, Bologna, Italy, October 2025, pp 1-9.
• Chao Feng, Alberto Huertas Celdran, Xi Cheng, Gérôme Bovet, Burkhard Stiller: GreenDFL: a Framework for Assessing the Sustainability of Decentralized Federated Learning Systems; arxiv, arxiv, Zürich, Switzerland, February 2025, URL

None