Abstract
Federated Learning (FL) faces significant challenges in wireless environments, specifically regarding data heterogeneity and communication bottlenecks. Standard FL algorithms often optimize for average performance, which can lead to poor results for agents with outlier data distributions. Furthermore, traditional communication protocols like TDMA and FDMA scale poorly as the number of participants increases. This paper introduces FedFAir, an algorithm that leverages Over-the-Air (OtA) computation and minmax optimization to achieve both fairness and robustness. By utilizing the physical properties of wireless signals for data aggregation, FedFAir eliminates the need for Channel State Information (CSI) and provides constant-time communication overhead, significantly outperforming standard methods in accuracy and efficiency.
Introduction
Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models across decentralized agents while keeping data localized. However, deploying FL in real-world wireless networks introduces two primary obstacles. First, data heterogeneity—where agents possess varying data distributions or noise levels—often results in a global model that performs well on average but fails for specific "outlier" agents. Second, communication bottlenecks arise because standard protocols require agents to transmit updates sequentially or on separate frequencies, leading to a linear increase in latency as the network grows.
To address these issues, this research proposes FedFAir, a framework designed to ensure fairness across heterogeneous agents while utilizing Over-the-Air (OtA) computation to maintain communication efficiency.
The FedFAir Framework
1. Minmax Optimization for Fairness
Traditional FL typically minimizes the empirical risk, which is the average loss across all agents. FedFAir shifts this focus by solving a minmax problem:
$$\min_{\theta} \max_{i} g_i(\theta)$$
By optimizing for the agent with the highest loss (the "worst-performing" agent), the algorithm ensures that the resulting model is robust to data skewness and noise. To solve this efficiently, the authors reformulate the problem into an epigraph form, introducing a dummy variable $\alpha$ that serves as an upper bound on the individual losses. This approach prevents the global model from ignoring minority data distributions or noisy sensors.
2. CSI-Free Over-the-Air Computation
OtA computation allows agents to transmit their signals simultaneously over the same frequency. The receiver (the central unit) naturally obtains the mathematical sum of these signals due to the superposition property of the wireless medium.
While OtA significantly reduces communication time, most existing methods require the central unit to know the exact Channel State Information (CSI) to decode the updates. FedFAir overcomes this by employing a three-round transmission protocol. By transmitting a constant value in a specific round, the receiver can normalize the aggregated signal without ever needing to know the individual channel gains. This "CSI-free" approach simplifies the hardware requirements and enhances privacy, as the central unit cannot "unmix" the signals to see individual agent contributions.
Experimental Evaluation
The authors evaluated FedFAir using a Federated Logistic Regression task involving 12 agents. To simulate real-world challenges, each agent was assigned a different number of training samples and unique noise distributions.
Performance and Accuracy
In these heterogeneous settings, FedFAir demonstrated a clear advantage over the standard FedAVG algorithm:
- Accuracy: FedFAir achieved approximately 90% accuracy after 5,000 iterations. In contrast, FedAVG struggled to reach 70–75% accuracy, even after 100,000 iterations.
- Class Imbalance: Confusion matrix analysis revealed that FedAVG often failed to identify specific classes (e.g., label '0') because it prioritized the majority distributions. FedFAir successfully recognized all classes by focusing on the worst-case performance.
Communication Efficiency
Because OtA computation allows for simultaneous transmission, the communication time for FedFAir remains constant regardless of the number of agents ($N$). For the 12-agent setup, FedFAir was 4 times faster per iteration than FedAVG. As the network scales to 30 or more agents, the speedup factor is expected to increase linearly (e.g., 10x faster), making it ideal for massive IoT deployments.
Conclusion
FedFAir represents a significant step forward in making Federated Learning viable for wireless sensor networks and IoT devices. By treating fairness as a form of robustness and leveraging the physics of the wireless channel, the algorithm provides a scalable, private, and high-performance solution for decentralized learning.
Future research will focus on extending these theoretical guarantees to non-convex deep learning models and enhancing the system's resilience against malicious "Byzantine" agents who might attempt to manipulate the minmax optimization process.
Summary updated