💻 computer science

Geometric Fairness-Aware Routing for Federated Edge Networks

This paper introduces Geo-FairFed, a geometric fairness-aware routing system for federated edge networks that leverages hyperbolic graph neural networks and curvature-regularized optimization to simultaneously minimize latency and energy consumption while significantly improving performance equity across distributed devices.

Original authors: Ratun Rahman

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Ratun Rahman

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The Traffic Jam Problem

Imagine a massive, sprawling city (the network) with millions of drivers (data packets) trying to get to their destinations. In the future (6G and Edge networks), these drivers are all different: some have fast sports cars (powerful servers), while others are on bicycles or walking (small, battery-powered IoT devices).

The Problem:
Current traffic systems (routing algorithms) are like a greedy GPS. They only care about getting the total number of cars to their destination as fast as possible. To do this, they send everyone down the wide, fast highways.

The Result: The sports cars zoom through, but the bicycles get stuck in tiny, slow side streets or are told to wait forever because the system ignores them. The "rich" nodes get all the speed, and the "poor" nodes suffer.

The Goal:
The authors want a system that is fair. They want the sports cars to move fast, but they also want to make sure the bicycles aren't left behind. They want to balance speed with equality.

The Solution: Geo-FairFed

The paper proposes a new system called Geo-FairFed. It combines three big ideas to solve the traffic jam fairly.

1. The Map: Hyperbolic Geometry (The "Funnel" Analogy)

Most computer maps are drawn on a flat piece of paper (Euclidean space). But real networks aren't flat; they are hierarchical. Think of a family tree or a corporate ladder: you have a few bosses at the top and thousands of employees at the bottom.

The Old Way: Trying to draw a giant family tree on a flat sheet of paper makes the bottom part squished and messy. Distances get distorted.
The New Way (Geo-FairFed): The authors use Hyperbolic Geometry. Imagine the map isn't flat paper, but the inside of a funnel or a saddle.
- In this "funnel world," the wide bottom has plenty of room for the thousands of small devices, while the narrow top holds the few big hubs.
- This shape naturally understands the "hierarchy" of the network. It helps the system see that some paths are naturally longer or more crowded than others, allowing it to route traffic more intelligently.

2. The Teamwork: Federated Learning (The "Secret Recipe" Analogy)

Usually, to fix traffic, a central computer (like a city mayor) needs to see every single car's location. This is slow and a privacy risk.

The New Way: The authors use Federated Learning. Imagine every driver has a notebook. Instead of sending their location to the mayor, they just write down their own "best route advice" in their notebook and send only the advice to a central aggregator.
The aggregator mixes all these tips together to create a "Master Recipe" for routing, which is then sent back to everyone. No one shares their private data, but everyone learns from the group.

3. The Fairness Rule: The "Level Playing Field" (The "Weighted Vote" Analogy)

Here is the tricky part. In standard teamwork, if one driver has a super-fast car and sends 1,000 tips, their advice might drown out the tips from the 1,000 bicycle riders. The system becomes biased toward the powerful.

The Fix: Geo-FairFed adds a Fairness Penalty.
- Imagine a voting system where the "rich" drivers (powerful nodes) have their votes slightly reduced if they are already doing too well, and the "poor" drivers (weak nodes) get a boost.
- The system constantly checks a "Fairness Score" (called Jain's Fairness Index). If the score drops (meaning some nodes are being treated unfairly), the system automatically adjusts the "Master Recipe" to help the struggling nodes, even if it slows the whole group down just a tiny bit.

How It Works in Practice

The system runs in a loop:

Local Learning: Each device looks at its own neighborhood using the "Funnel Map" (Hyperbolic space) to figure out the best local route.
Sharing: Devices send their "route advice" (model updates) to the central server.
Fair Mixing: The server mixes the advice. But it doesn't just average them. It uses a special formula that says, "If Node A is doing great and Node B is struggling, let's give Node B's advice a little more weight."
Result: A new, fairer global map is sent back to everyone.

The Results (What the Paper Found)

The authors tested this on simulated 6G and Internet networks. They compared Geo-FairFed against the best existing methods.

Faster: The average time it took for data to travel dropped by 20%.
Greener: The energy used by the devices dropped by 17%.
Fairer: The "Fairness Score" improved by up to 21%.

The Key Takeaway:
By using a "funnel-shaped" map (Hyperbolic geometry) to understand the network's shape and adding a "fairness rule" to the teamwork process, the system managed to make the whole network faster and more efficient without leaving the smaller, weaker devices behind. It proved you can have a fast network that is also a kind network.

Technical Summary: Geometric Fairness-Aware Routing for Federated Edge Networks (Geo-FairFed)

Problem Statement
The paper addresses the critical challenge of routing in emerging 6G and edge-intelligent networks, which consist of billions of diverse, geographically distributed devices. Existing routing systems face two primary limitations:

Lack of Fairness: Traditional optimization frameworks often prioritize aggregate metrics (latency, throughput) at the expense of individual node equity. Resource-rich nodes tend to dominate routing decisions, leaving weaker or remote devices with degraded service.
Geometric Mismatch: Most current machine learning-based routing models, including those using Graph Neural Networks (GNNs), operate in Euclidean embedding spaces. These spaces fail to accurately capture the hierarchical, power-law, and asymmetric nature of real-world network topologies, leading to suboptimal traffic load balancing.

Furthermore, while Federated Learning (FL) offers privacy-preserving distributed training, standard FL aggregation methods (like FedAvg) often assume statistical homogeneity among clients, resulting in biased global models in heterogeneous edge environments.

Methodology: The Geo-FairFed Framework
The authors propose Geo-FairFed, a federated routing architecture that integrates hyperbolic geometric learning with fairness-constrained optimization. The framework operates through three iterative stages:

Local Geometric Learning (HGNN):
Each edge node trains a local routing policy using a Hyperbolic Graph Neural Network (HGNN). Instead of Euclidean space, node embeddings are mapped to a negatively curved hyperbolic manifold ( $H_d$ ). This allows the model to encode hierarchical relationships and connection asymmetries inherent in network topologies. The local objective minimizes a composite loss function comprising end-to-end latency, energy consumption, and packet drop rates, alongside a geometric consistency loss to maintain manifold integrity.
Fairness-Aware Aggregation:
To address the bias in standard federated aggregation, Geo-FairFed introduces a fairness-adjusted weighting mechanism. Instead of weighting updates solely by sample size (as in FedAvg), the aggregator calculates a weight based on a node's local performance inequality.
- The system utilizes Jain's Fairness Index to quantify local performance equity.
- Nodes exhibiting higher inequality (lower fairness scores) receive lower aggregation weights, preventing them from disproportionately influencing the global model if they are dominating resources, while ensuring underperforming nodes are not ignored.
- The global objective function minimizes a weighted sum of routing loss, a fairness penalty (derived from Jain's index), and geometric distortion.
Joint Optimization:
The framework employs a unified optimization strategy that balances routing efficiency, geometric consistency, and fairness. Theoretical analysis establishes convergence guarantees under assumptions of smooth local loss functions, bounded gradient variance, and limited manifold curvature.

Key Contributions

Hyperbolic Representation for Routing: The introduction of HGNNs to learn topology-aware representations in a negatively curved manifold, enabling efficient encoding of hierarchical and asymmetric patterns common in large-scale edge networks.
Fairness-Constrained Federated Aggregation: The development of a federated routing method that integrates Jain's Fairness Index directly into the aggregation process, dynamically adjusting weights to balance performance across heterogeneous devices.
Unified Optimization and Theoretical Guarantees: A proposed optimization strategy that minimizes routing delay, geometric inconsistency, and injustice simultaneously. The paper provides a theoretical proof showing convergence under limited curvature and demonstrates that the fairness term leads to a Pareto-improving equilibrium.
Comprehensive Evaluation: Extensive simulations on dynamic 6G and IoT topologies (including synthetic Barabási-Albert graphs and real-world ISP datasets like RocketFuel and TopologyZoo) validating the framework's efficacy.

Experimental Results
Simulations demonstrate that Geo-FairFed outperforms state-of-the-art federated and geometric routing protocols (including DQN-Routing, FedAvg-Routing, and standard HGNN-Routing). Key performance metrics include:

Latency: A reduction in average latency by approximately 20%.
Energy Consumption: A decrease in energy usage by 17%.
Fairness: An improvement in the Jain's Fairness Index by up to 21%.
Convergence: The framework achieves faster convergence to a lower steady-state loss and a higher fairness plateau compared to baselines.
Curvature Sensitivity: Experiments indicate that a modestly hyperbolic curvature (e.g., $\kappa \approx -0.3$ ) yields optimal routing accuracy, confirming that hyperbolic geometry better represents network hierarchies than Euclidean space ( $\kappa=0$ ).

Significance and Claims
The paper claims that Geo-FairFed significantly enhances both the efficiency and equity of large-scale network routing. By embedding topology in a hyperbolic manifold and integrating fairness into federated updates, the system achieves a Pareto-improving equilibrium, meaning it improves fairness without causing significant degradation in overall routing efficiency.

The authors emphasize that this approach bridges the gap between geometric representation learning and distributed network optimization. They assert that the framework is particularly effective in heterogeneous 6G and edge environments where traditional Euclidean models and standard federated averaging fail to account for topological hierarchies and node disparities. The study concludes that incorporating geometric fairness-awareness is essential for the next generation of scalable, intelligent communication networks.

Note: The paper acknowledges limitations regarding real-world deployment factors such as stragglers, unstable connections, and adversarial attacks, noting that these are areas for future research rather than current capabilities of the proposed framework.