Hierarchical Federated Learning for 6G Edge Computing
| Reference number | |
| Coordinator | Uppsala universitet - Institutionen för elektroteknik |
| Funding from Vinnova | SEK 1 784 674 |
| Project duration | October 2024 - October 2026 |
| Status | Ongoing |
| Venture | 6G - Research and innovation |
| Call | 6G International research and innovation cooperation 2024 |
Important results from the project
We developed a flexible framework to test hierarchical federated learning (HFL) under diverse device and network conditions. We developed resource-aware training and aggregation methods using layer freezing and Switch-Before-Aggregate (SBA), which improved model accuracy over the tested FL methods. Full HFL integration remains a next step. Spin-off results include studies on resource-efficient LLM training, compression and serving, and DynaStitch for ViT inference, presented at EMSOFT 2026.
Expected long term effects
Our work could help devices with diverse computing resources, such as phones, connected vehicles, and other edge devices, participate in collaborative AI training in future 6G environments. We will extend SBA to hierarchical FL and multi-task learning, and pursue more communication-efficient and privacy-aware model exchange. Strong ties with academic and industrial partners across Sweden, Korea, and Germany provide a solid foundation for joint research and industry-relevant innovation.
Approach and implementation
We planned an HFL testbed, parallel development of three techniques, and integration. We built the testbed and successfully developed core resource-aware training and aggregation methods, including SBA. We adapted the testbed for practical validation, following our risk plan. Feedback from international academic and industry partners refined our approach. Delayed postdoc recruitment led to a six-month extension, but the research progressed broadly as planned on the revised schedule.