Abstract
Federated Learning (FL) typically faces significant communication bottlenecks when aggregating model updates from numerous edge devices. Over-the-Air (OtA) computation offers a bandwidth-efficient alternative by leveraging the superposition property of wireless channels. However, existing OtA-FL schemes often rely on perfect Channel State Information at the Transmitter (CSIT) and assume idealized analog transmission. This paper introduces FedCPU (Compute-Update FL), a universal framework that utilizes lattice joint source-channel coding to perform "blind" digital OtA aggregation. FedCPU eliminates the need for CSIT and provides a robust "safety bubble" against wireless noise through lattice-based quantization. Our evaluations show that FedCPU achieves near-ideal performance in challenging fading environments, significantly outperforming prior blind aggregation benchmarks.
Introduction
The proliferation of Internet of Things (IoT) devices has fueled the demand for Federated Learning (FL), where models are trained locally on edge devices and aggregated at a central server. Traditional orthogonal communication methods, such as Time Division Multiple Access (TDMA), scale poorly with the number of devices. Over-the-Air (OtA) computation addresses this by allowing simultaneous transmissions, where the channel naturally performs the summation required for model averaging.
Despite its potential, OtA-FL faces two primary hurdles: the requirement for precise CSIT to align signals and the discrepancy between theoretical analog models and practical digital hardware. FedCPU addresses these by treating model quantization and wireless interference as a joint coding problem using lattice structures.
The FedCPU Framework
The core innovation of FedCPU lies in its use of Lattice Joint Source-Channel Coding (JSCC). Instead of separate source and channel coding, the lattice structure simultaneously handles compression and error protection.
Transmitter Design
Each device $k$ processes its model update $\Delta \mathbf{w}_k$ through a sequence of operations: normalization, addition of a random dither $\mathbf{d}_k$, and lattice quantization $Q_\Lambda$. The dither ensures that quantization errors are statistically independent of the signal, preventing bias in the global model.
Receiver and Aggregation
The server employs a two-layer architecture. An equalization layer uses multiple antennas to minimize channel-induced errors, while a normalization layer compensates for quantization noise. Crucially, the server decodes an integer combination of lattice points. Because the sum of lattice points is itself a lattice point, the aggregation process is inherently robust to interference, provided the noise stays within the lattice's "Voronoi region" or safety bubble.
Experimental Evaluation
We evaluated FedCPU using the MNIST dataset in a non-i.i.d. setting with 30 devices and a 30-antenna server. The wireless environment was modeled using Rayleigh fading with Gaussian noise.
Performance Comparison
FedCPU was compared against two benchmarks: BFEL (Blind Federated Edge Learning), a state-of-the-art method without quantization, and Ideal UVeQFed, a theoretical upper bound assuming error-free orthogonal communication.
As shown in the results, FedCPU significantly outperforms BFEL and closely tracks the performance of the ideal scenario. The lattice structure provides a level of resilience that allows the learning process to remain stable even in the absence of CSIT.
Key Insights
- Complexity Shift: By removing the CSIT requirement, FedCPU shifts the computational burden from power-constrained edge devices to the server, which can leverage multiple antennas to untangle signals.
- Lattice Density Trade-off: The system performance depends on finding a "Goldilocks" zone for lattice density—balancing quantization precision against noise robustness.
- Scalability: Accuracy improves with the number of server antennas ($M$), though with diminishing returns, indicating the scheme's efficiency even with moderate hardware configurations.
Conclusion
FedCPU represents a significant step toward practical, "blind" over-the-air federated learning. By integrating lattice coding into the FL pipeline, the framework bridges the gap between theoretical OtA computation and real-world digital wireless systems. Future work will focus on scaling the approach to more complex datasets like CIFAR-10 and exploring "doubly blind" scenarios where neither the transmitter nor the receiver has perfect channel knowledge.
Summary updated