Introduction
In wireless Federated Learning (FL), a central Parameter Server (PS) coordinates with distributed devices to train a global model. A significant challenge in practical deployments is the "coherence disparity" between devices with different mobility profiles. For instance, a static smart appliance and a high-speed mobile device experience vastly different wireless channel stability. Conventional FL often employs a "one-size-fits-all" communication strategy, which leads to bandwidth waste for static devices or corrupted updates for mobile ones. This paper introduces a coherence-aware framework that leverages this disparity to improve communication efficiency.
System Model and Problem Formulation
The proposed system considers a set of static devices ($\mathcal{K}_S$) and dynamic devices ($\mathcal{K}_D$). Static devices have long coherence intervals, while dynamic devices require frequent pilot signals for Channel State Information (CSI) estimation.
The core problem is the inefficient use of pilot signals. In high-mobility scenarios, a large portion of the downlink resources is dedicated to pilots, leaving less room for model parameter transmission. This paper addresses this by proposing a resource-reuse strategy.
Proposed Methodology: Product Superposition
The authors propose product superposition, a technique that overlays global model updates onto the pilot signals. Instead of keeping pilots and data orthogonal, the PS embeds partial model updates within the pilot symbols.
Dual-Purpose Signaling
This approach serves two purposes simultaneously:
- For Static Devices: Since these devices have high channel coherence and already possess accurate CSI, they can directly decode the model updates embedded in the pilot slots.
- For Dynamic Devices: These devices use the received superposition signal to estimate a "virtual channel"—the product of the physical channel and the embedded parameters. This virtual channel estimate then facilitates the decoding of the remaining model data in standard data slots.
Handling Partial Updates
Dynamic devices may only receive a subset of the model parameters due to the superposition. The authors evaluated two strategies for filling these gaps: Zero-Filling (ZF) and Previous Local Model Filling (PLMF). PLMF, which utilizes parameters from the previous training round, significantly outperforms ZF by reducing gradient bias.
Experimental Evaluation
The framework was tested on MNIST (with a CNN) and CIFAR-10 (with ResNet-18) datasets, considering both i.i.d. and non-i.i.d. data distributions.
Key Results
- Efficiency Gains: On the MNIST dataset, the proposed scheme reduced the normalized communication cost by approximately 0.3 to reach 95% test accuracy compared to traditional orthogonal signaling.
- Robustness: As shown in the results, the PLMF-based scheme maintains high accuracy even as the pilot overhead ($\lambda$) increases, whereas conventional methods and ZF show significant performance degradation.
- Scalability: For the more complex CIFAR-10 dataset, the framework achieved a 0.28 reduction in communication cost at 66% accuracy (SNR = 30dB).
Conclusion and Future Work
This research demonstrates that coherence disparity, typically viewed as a hindrance, can be exploited to enhance FL efficiency through product superposition. By embedding data within pilots, the framework significantly reduces communication overhead without sacrificing convergence. Future research directions include extending this coherence-aware approach to the uplink and exploring joint power allocation strategies for both downlink and uplink transmissions.
Summary updated