💻 computer science

Federated Learning Meets Random Access: Energy-Efficient Uplink Resource Allocation

This paper proposes energy-efficient uplink resource allocation strategies for wireless networks hosting concurrent federated learning and random access traffic, demonstrating that ALOHA outperforms slotted-ALOHA in FL-dominated scenarios while slotted-ALOHA is superior when random access traffic prevails.

Original authors: Giovanni Perin, Eunjeong Jeong, Nikolaos Pappas

Published 2026-08-26✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Giovanni Perin, Eunjeong Jeong, Nikolaos Pappas

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Wireless networks have long been designed to carry information mostly in one direction: from a central tower down to our phones and devices. This pattern, where downloads vastly outnumber uploads, has shaped how engineers build the invisible infrastructure of the internet. However, a new wave of artificial intelligence is rewriting these rules. Modern AI applications, particularly those that generate content or learn from data, require devices to send massive amounts of information back to the network. This shift creates a bottleneck, as the same limited radio waves must now carry both the familiar, steady stream of downloads and these new, heavy bursts of data going up. The challenge is no longer just about speed; it is about how to share a crowded highway when two very different types of traffic need to move at the same time.

In a recent study, researchers Giovanni Perin, Eunjeong Jeong, and Nikolaos Pappas tackled this specific problem of sharing wireless resources. They focused on a scenario where two distinct groups of devices are trying to communicate with a central gateway using the same slice of radio spectrum. One group consists of machines working together to train an artificial intelligence model. These devices operate in a highly organized way, taking turns to upload large chunks of data so the central system can learn. The other group represents devices that need to send frequent, smaller updates, such as sensor readings or AI inference results. These devices do not take turns; instead, they try to send data whenever they have something to say, a method known as random access. The researchers wanted to find the most energy-efficient way to split the available bandwidth between these two groups while ensuring the AI training finishes on time and the random-access devices get enough data through.

The team built a mathematical model to simulate this interaction, treating the bandwidth as a pie that could be sliced in different proportions. They tested two different strategies for the random-access devices: one where they simply shout out their data whenever they want, and another where they wait for synchronized time slots to speak. The goal was to minimize the total electricity used by the entire system. Their simulations revealed that there is no single "best" strategy that works for every situation; the optimal choice depends entirely on which group is using more of the network's capacity.

When the system is dominated by the heavy data uploads of the AI training devices, the simple, uncoordinated shouting method proves surprisingly effective. In these cases, allowing the random-access devices to transmit without strict timing rules actually saves significant energy, reducing consumption by nearly half compared to the more organized approach. This is because the simple method allows the AI devices to claim a larger share of the bandwidth, letting them finish their massive uploads faster and use less power overall. However, the situation flips when the random-access traffic becomes the primary load. If the network is flooded with frequent small updates, the synchronized, time-slot method becomes the clear winner. In these crowded conditions, the organized approach reduces collisions and retransmissions, leading to a modest but meaningful drop in energy use, about six percent lower than the uncoordinated method.

The researchers also discovered that the number of devices involved changes the outcome. When there are fewer AI devices, the simple random method remains efficient even as the number of random-access devices grows. But as the number of AI devices increases, the network becomes more sensitive to how the bandwidth is split. If the demand for random access is high, the synchronized method becomes necessary to keep the system stable, even though it forces the AI devices to wait longer and use more energy. The study suggests that network managers should not stick to one protocol forever. Instead, they should monitor the traffic mix. If the network is busy with heavy AI training, the uncoordinated approach is the most efficient path. If the network is instead clogged with a flood of small, frequent messages, switching to the synchronized, time-slot approach will save energy and keep the system running smoothly.

This work highlights a fundamental shift in how we must think about wireless networks. The old assumption that one protocol fits all is no longer valid in an era where artificial intelligence and traditional data streams compete for the same space. By carefully adjusting how much bandwidth is given to each type of traffic and choosing the right communication style for the moment, networks can operate with much greater efficiency. The findings do not promise a magic solution that eliminates all energy costs, but they provide a clear, practical guide for balancing these competing demands. The key takeaway is that flexibility is the most valuable resource of all; knowing when to let devices speak freely and when to make them wait in line can make the difference between a network that struggles and one that thrives.

Technical Summary: Federated Learning Meets Random Access: Energy-Efficient Uplink Resource Allocation

Problem Statement
The paper addresses the emerging challenge of resource scarcity in wireless networks caused by the coexistence of traditional communication tasks and bandwidth-intensive Artificial Intelligence (AI) applications. Specifically, it focuses on the uplink resource allocation problem where two distinct sets of devices share the same bandwidth:

Federated Learning (FL) Devices: A set of $N$ devices performing distributed model training. These devices transmit large model updates periodically using Frequency Division Multiple Access (FDMA). Their primary constraint is a strict latency budget for the learning round.
Throughput-Oriented Devices: A set of $K$ devices generating frequent, bandwidth-intensive traffic (e.g., GenAI inference streams or IoT updates). These devices access the medium via Random Access (RA) protocols, specifically comparing Pure ALOHA and Slotted ALOHA (S-ALOHA). Their primary requirement is a guaranteed minimum throughput.

The core conflict arises because FL traffic requires large, scheduled bandwidth allocations to meet latency constraints, while RA traffic competes for the remaining spectrum. The authors identify a gap in existing literature, which often treats traffic as homogeneous (e.g., URLLC vs. eMBB) or focuses on joint optimization of sensing/computation without addressing the specific medium access conflicts between FDMA-based FL and RA-based throughput services.

Methodology
The authors formulate a joint radio resource allocation problem aimed at minimizing the global system energy consumption. The decision variables are:

$\rho$ : The bandwidth share allocated to FL devices (with $1-\rho$ allocated to RA devices).
$\lambda$ : The total (re)transmission rate of the RA devices.

The optimization is subject to four constraints:

FL Latency: The total transmission time for FL updates must fit within the remaining time of the learning round after computation and idling.
RA Throughput: The RA devices must achieve a minimum guaranteed throughput $q$ .
Bandwidth Bounds: $\rho$ must be between 0 and 1.
System Stability: A minimum retransmission rate $\varepsilon$ is enforced to ensure stability.

The problem is non-convex. To solve it, the authors employ a decomposition approach:

Optimization of $\lambda$ : For a fixed bandwidth share $\rho$ , the optimal transmission rate $\lambda^*(\rho)$ is derived. Since throughput is monotonic in $\lambda$ , the optimal energy solution lies at the boundary of the throughput constraint (minimizing retransmissions).
Optimization of $\rho$ : With $\lambda^*(\rho)$ determined, the authors perform a grid search over the feasible range of $\rho$ to find the global minimum of the total energy function $E(\lambda^*(\rho), \rho)$ .

The energy models account for Shannon capacity limits, channel gains, noise, and the specific collision probabilities inherent to ALOHA and S-ALOHA under varying bandwidth allocations.

Key Results
Numerical experiments, utilizing parameters such as 30 FL devices, a 60 MHz bandwidth, and varying RA arrival rates, yield the following insights:

Protocol Performance Trade-offs:
- ALOHA: Demonstrates superior flexibility when FL traffic dominates the system. It maintains near-optimal throughput even when allocated high bandwidth shares ( $\rho$ ), allowing FL devices to transmit faster and with lower energy. In specific configurations where the system is dominated by FL traffic (e.g., Configuration C1), ALOHA yields 48.8% lower energy consumption compared to S-ALOHA.
- S-ALOHA: Becomes more efficient when RA traffic volume is high. While its throughput decreases linearly as $\rho$ increases (due to the vulnerable window constraint), it offers higher success probabilities for RA packets in high-load scenarios. When RA traffic dominates, S-ALOHA yields approximately 6% lower energy consumption.
Sensitivity Analysis:
- As the number of FL devices ( $N$ ) increases, the system energy consumption rises. ALOHA remains the preferred protocol up to a threshold (approx. $N=60$ in the tested configuration), beyond which S-ALOHA becomes more efficient.
- Increasing the RA arrival rate ( $\lambda'$ ) degrades ALOHA's efficiency more rapidly than S-ALOHA's.
Configuration Case Studies:
- In configurations with low RA traffic (e.g., $N=10$ , low $\lambda'$ ), ALOHA significantly outperforms S-ALOHA in energy efficiency.
- In configurations with high RA traffic or long FL round durations (e.g., $N=30$ , high $\lambda'$ ), S-ALOHA becomes the more energy-efficient choice, though the margin is smaller.

Significance and Claims
The paper provides a close-to-optimal solution to the specific non-convex optimization problem defined by the authors, with the reported grid-search resolution of 1 percent bounding the optimality gap. The primary contribution is the demonstration that an adaptive approach, which switches between ALOHA and S-ALOHA based on the traffic mix and system parameters, could be highly beneficial for energy efficiency.

The authors modestly conclude that neither protocol universally outperforms the other. Instead, the optimal choice depends on the dominance of traffic types:

ALOHA is preferable when the system energy is dominated by FL devices (large model uploads), as it allows for higher bandwidth shares to FL, reducing transmission time and energy.
S-ALOHA is preferable when RA traffic dominates, as it supports higher throughputs and lower collision rates under heavy load, thereby reducing the energy cost of retransmissions for the RA devices.

The work highlights that as AI-generated traffic (particularly uplink-heavy GenAI and FL) reshapes wireless networks, static resource allocation strategies are insufficient; dynamic, traffic-aware allocation is critical for energy efficiency.

Technical Summary: Federated Learning Meets Random Access: Energy-Efficient Uplink Resource Allocation

More like this