Abstract
Federated Learning (FL) has emerged as a privacy-preserving paradigm for training machine learning models across distributed devices. However, in practical scenarios, multiple FL services often compete for limited data and communication resources. This paper addresses the challenge of resource allocation and pricing in Unmanned Aerial Vehicle (UAV)-aided networks. We propose a tripartite market model involving FL service demanders, data-owning devices, and UAV aggregators. By introducing a "joint bid" mechanism that pairs data-sellers with UAV-sellers, we simplify the complex matching problem. We develop two auction mechanisms—an optimal VCG-based approach and a computationally efficient suboptimal matching algorithm—to ensure truthfulness, individual rationality, and high revenue for the service platform.
Introduction
Federated Learning (FL) allows smart devices to train models locally, sharing only model updates rather than raw data. While effective, the efficiency of FL is often hampered by the distance between devices and central cloud servers, leading to high latency and energy consumption. Unmanned Aerial Vehicles (UAVs) can mitigate this by acting as mobile edge aggregators, providing localized model aggregation and relay services.
In a real-world ecosystem, multiple FL Service Demanders (FLSDs) may simultaneously seek data from the same pool of devices. This creates a complex competition for both data and UAV resources. To manage this, a robust trading mechanism is required to match buyers (FLSDs) with sellers (data owners and UAVs) while ensuring economic fairness and computational feasibility.
System Architecture and the Tripartite Market
The proposed system operates as a tripartite market consisting of three primary stakeholders:
- Buyers (FLSDs): Entities requiring trained models for specific tasks.
- Data-Sellers: Smart devices or groups with local datasets.
- UAV-Sellers: UAVs that provide the necessary flight and processing power for model aggregation.
The interaction follows a seven-step process: (1) FLSDs submit service requirements; (2) the platform initiates an auction; (3) devices perform local training; (4) models are uploaded to UAVs; (5) UAVs aggregate the models; (6) trained models are submitted to the platform; and (7) the final service is delivered to the buyers.
Auction Mechanisms and Seller Pairing
The core innovation of this work is the virtualization of the tripartite matching problem. Matching three independent parties is mathematically complex (hypergraph matching). To simplify this, the authors introduce seller pairing and joint bidding. A data-seller and a UAV-seller are combined into a single "virtual seller pair" that submits a joint bid representing their combined costs.
Two distinct reverse auction mechanisms are proposed:
- VCG-Based Optimal Auction: Utilizing the Vickrey-Clarke-Groves mechanism, this approach maximizes total buyer revenue. It guarantees truthfulness (sellers cannot profit by lying about costs) and individual rationality (all participants make a profit). However, it is computationally intensive (NP-hard).
- One-Sided Matching Suboptimal Auction: To ensure practical deployment in dynamic networks, this polynomial-time algorithm uses preference lists and "critical value" payment rules. It maintains essential economic properties while significantly reducing computational overhead.
Performance Evaluation
The mechanisms were evaluated using numerical simulations based on the MNIST dataset. Key findings include:
- Computational Efficiency: In a scenario with 9 buyers and 10 sellers, the optimal VCG algorithm required over 25 days to compute. In contrast, the proposed suboptimal matching algorithm achieved a solution in just 1.3 seconds.
- Revenue Maximization: The suboptimal algorithm achieved a total buyer revenue within 2.25% of the theoretical maximum, demonstrating that speed does not come at a significant cost to performance.
- Economic Integrity: Simulations confirmed that the mechanisms are truthful; no seller could increase their utility by misreporting their true costs. This stability is crucial for maintaining a healthy trading ecosystem.
Conclusion and Future Work
This paper provides a scalable and economically sound framework for trading FL services in UAV-aided networks. By simplifying the market through seller pairing, the proposed mechanisms offer a balance between theoretical optimality and practical execution. Future research will explore multi-source scenarios where a single FL task can aggregate data from multiple subregions simultaneously, as well as dynamic auction models that account for real-time UAV mobility and battery constraints.
Summary updated