Abstract
We propose an improved convergence analysis technique that characterizes the distributed learning paradigm of federated learning (FL) with imperfect/noisy uplink and downlink communications. Such imperfect communication scenarios arise in the practical deployment of FL in emerging communication systems and protocols. The analysis developed in this paper demonstrates, for the first time, that there is an asymmetry in the detrimental effects of uplink and downlink communications in FL. In particular, the adverse effect of the downlink noise is more severe on the convergence of FL algorithms. Using this insight, we propose improved Signal-to-Noise (SNR) control strategies that, discarding the negligible higher-order terms, lead to a similar convergence rate for FL as in the case of a perfect, noise-free communication channel while incurring significantly less power resources compared to existing solutions. In particular, we establish that to maintain the $O(1/\sqrt{K})$ rate of convergence like in the case of noise-free FL, we need to scale down the uplink and downlink noise by $\Omega(\sqrt{k})$ and $\Omega(k)$ respectively, where $k$ denotes the communication round. Our theoretical result is further characterized by two major benefits: firstly, it does not assume the somewhat unrealistic assumption of bounded client dissimilarity, and secondly, it only requires smooth non-convex loss functions, a function class better suited for modern machine learning and deep learning models.
Introduction
Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models across distributed edge devices while keeping data localized. However, the transition from theoretical models to real-world deployments—such as IoT sensors and mobile devices—introduces a significant hurdle: imperfect communication channels. While most standard FL research assumes "perfect" communication, practical channels are inherently noisy and unreliable.
Existing literature addressing noisy FL often relies on restrictive assumptions, such as Bounded Client Dissimilarity (BCD) or strong convexity, which rarely hold true for modern deep learning tasks. Furthermore, previous strategies for mitigating noise often suggested aggressive power scaling for both uplink and downlink communications, leading to inefficient energy consumption on battery-constrained edge devices. This paper addresses these gaps by providing a tighter mathematical analysis for non-convex neural networks and proposing a power-efficient Signal-to-Noise Ratio (SNR) control strategy.
The Asymmetry of Noise in Federated Learning
The core contribution of this research is the discovery of a critical asymmetry between the two directions of communication in FL:
- The Downlink Bottleneck (Server-to-Client): In the downlink, the server broadcasts a single global model to all participating clients. If this model is corrupted by noise, every client begins their local training from a flawed starting point. This error compounds over successive communication rounds, creating a persistent error floor ($O(1)$) that prevents the model from converging to the global optimum.
- Uplink Resilience (Client-to-Server): In the uplink, multiple clients send their local updates back to the server. Because the server averages these updates, the random, zero-mean noise from individual clients tends to cancel out. Consequently, uplink noise is "transient" and contributes a less severe $O(1/\sqrt{K})$ error term.
Proposed Method: Noisy-FedAvg and Asymmetric SNR Scaling
To address this asymmetry, the authors introduce Noisy-FedAvg, an extension of the standard Federated Averaging algorithm that explicitly accounts for additive noise in both directions. By analyzing the convergence properties of smooth non-convex functions without the BCD assumption, the authors derived a specific scaling policy to maintain the standard $O(1/\sqrt{K})$ convergence rate found in noise-free systems.
To achieve this "gold standard" convergence, the SNR must be controlled such that:
- Downlink noise is scaled down by $\Omega(k)$.
- Uplink noise is scaled down by only $\Omega(\sqrt{k})$.
This asymmetric approach allows the noise to become a "higher-order term" that vanishes faster than the learning error itself. Crucially, because the uplink requires less aggressive noise suppression, edge devices can significantly reduce their transmission power, extending battery life without sacrificing model accuracy.
Empirical Validation
The theoretical findings were validated across several benchmarks, ranging from synthetic linear regression to complex deep learning tasks using CNNs on the following datasets:
- MNIST & Fashion-MNIST: Standard image classification tasks.
- CIFAR-10: A more complex object recognition dataset.
- FEMNIST: A dedicated federated benchmark characterized by high heterogeneity (non-IID data).
In all experiments, models using "Constant Noise" (no SNR control) either diverged or plateaued at high loss levels. However, when the proposed asymmetric SNR control strategy was applied, the training curves became nearly indistinguishable from the "Noise-Free" (perfect communication) baseline. This held true even for the highly heterogeneous FEMNIST dataset, proving the robustness of the analysis in non-IID settings.
Conclusion and Future Work
This work provides a more realistic and efficient framework for deploying FL in noisy environments. By identifying that the downlink is the primary bottleneck, the authors have enabled a strategy that preserves convergence while optimizing power usage.
Future research directions include developing parameter-free versions of this scheme (removing the need to know the "smoothness" constant $L$ of the loss function) and investigating the absolute theoretical lower bounds of performance in noisy environments. Additionally, testing these strategies on specific radio hardware (e.g., 5G or WiFi) will be essential to account for real-world variables like fading and interference.
Summary updated