Turning Interference into Intelligence: Energy-Efficient Federated Learning for 6G Networks
Future communication networks are evolving beyond connectivity to enable connected intelligence, where communication, sensing and computation operate seamlessly together. As 6G systems move towards large-scale IoT deployments, one of the biggest challenges is ensuring that intelligent devices can learn and collaborate under strict energy and spectrum constraints.
In new research from Durham University and The University of Manchester, supported by the CHEDDAR Hub, researchers explore a novel approach to Analog Over-the-Air Federated Learning (OTA-FL) combined with interference-based energy harvesting, offering a pathway towards more scalable and sustainable intelligent networks.
Download Research Poster
The Challenge: Scaling AI in Energy-Constrained Networks
Federated learning enables distributed devices to collaboratively train AI models without sharing raw data. However, in wireless IoT environments, this approach faces several limitations:
- Dependence on channel state information (CSI) and power-intensive channel inversion
- Unpredictable energy availability in energy-harvesting devices
- Interference (CCI) degrading communication quality
- Fixed training workloads leading to device dropouts and reduced accuracy
These challenges limit the ability to scale intelligent systems across dense, low-power IoT networks.
The Innovation: Learning Through the Air – Powered by Interference
This research introduces a new system that combines:
- Analog Over-the-Air Aggregation
Devices transmit updates simultaneously, allowing the network to compute directly over the wireless channel reducing communication overhead.
- RF Energy Harvesting from Interference
Rather than treating interference as purely harmful, the system repurposes it as a source of energy, enabling self-sustaining devices.
- CSI-Free Denoising
A novel denoising method removes the need for channel state information, significantly reducing system complexity and energy consumption.
Adaptive Intelligence: Training That Responds to Energy Availability
A key contribution of this work is an energy-aware adaptive training mechanism, where each device dynamically adjusts:
- The number of local training epochs
- The size of the dataset used
This ensures that devices continue participating even under fluctuating energy conditions, improving reliability and overall model performance.
Results: More Efficient, More Scalable, More Sustainable
Simulation results demonstrate that the proposed approach:
- Achieves accuracy comparable to traditional CSI-based methods
- Converges faster under varying energy conditions
- Increases device participation
- Achieves target performance with lower total energy consumption
Why It Matters for 6G
This work highlights a fundamental shift in how future networks can operate:
- From energy-limited to energy-aware systems
- From interference mitigation to interference utilisation
- From rigid training to adaptive, resilient AI systems
Together, these innovations support the development of cost-effective, scalable, and sustainable 6G IoT ecosystems, aligning with CHEDDAR’s vision for AI-native and energy-efficient networks.
