Decentralized Edge Intelligence: A Dynamic Resource Allocation Framework for Hierarchical Federated Learning

Wei Yang Bryan Lim, Jer Shyuan Ng, Zehui Xiong, Jiangming Jin, Yang Zhang, Dusit Niyato, Cyril Leung, Chunyan Miao

Abstract

This paper proposes a two-level decentralized framework to optimize Hierarchical Federated Learning (HFL). By integrating evolutionary game theory to model worker-cluster selection and a deep learning-based auction to price and allocate cluster services, the framework ensures system stability, incentive compatibility, and maximized revenue for edge providers.

The Challenge of Edge Intelligence

Federated Learning (FL) enables AI model training on local data without compromising privacy. However, coordinating thousands of heterogeneous devices (e.g., smartphones, IoT sensors) directly with a central cloud server often leads to high latency and frequent dropouts.

Hierarchical Federated Learning (HFL) addresses this by introducing "cluster heads" (edge servers or base stations) for intermediate data aggregation. Yet, HFL introduces its own coordination hurdles:

  1. Worker Choice: How do workers decide which cluster to join when rewards and network congestion vary?
  2. Incentive Design: How can cluster heads and workers be compensated fairly while providing model owners with optimal resources?
  3. Dynamic Rationality: Real-world network participants adapt slowly and have limited information, unlike the "perfectly rational" agents assumed in traditional models.

The Two-Level Framework

To address these challenges, the authors propose a two-layer optimization strategy that balances the needs of workers, cluster heads, and model owners.

The proposed two-level framework, integrating evolutionary games for worker selection and deep learning auctions for resource allocation.

Lower Level: Evolutionary Game Theory

Instead of assuming workers are perfect optimizers, the framework utilizes replicator dynamics. This models a population of workers who migrate between clusters based on their net utility—calculated as rewards minus costs like computation and network congestion. Over time, the system reaches an Evolutionary Equilibrium where no worker has an incentive to switch, providing a stable foundation for training.

Upper Level: Deep Learning-Based Auction

Model owners bid for the services of cluster heads. Traditional auctions often struggle to balance "truthfulness" (bidders revealing their actual value) with "revenue maximization." The authors employ a neural network to learn monotone transform functions for bids.

The deep learning-based auction architecture, incorporating monotone transform functions to ensure incentive compatibility.

This architecture ensures the auction is Incentive Compatible (truthful) and Individually Rational while maximizing the cluster head's revenue. By solving for an optimal auction through loss function optimization rather than rigid manual formulas, the system adapts to complex economic constraints.

Key Insights and Results

  • Bounded Rationality is Realistic: By using evolutionary games, the model accounts for the fact that workers in massive IoT networks adapt gradually by observing others rather than calculating global optima instantly.
  • The Congestion-Reward Trade-off: The paper mathematically proves that a stable equilibrium always exists between the attraction of high reward pools and the deterrent of communication congestion.
  • Revenue Superiority: Simulations show that the deep learning-based auction consistently generates higher revenue for cluster heads compared to standard Second Price Auctions (SPA), while maintaining truthfulness.
  • Impact of Data Coverage: Clusters with higher data coverage (more workers or higher quality data) were valued significantly higher. For instance, a cluster with 0.4 data coverage earned approximately 25% more revenue than one with 0.26 coverage when model owners required high accuracy.

Conclusion and Future Directions

The proposed framework demonstrates that combining game-theoretic stability with deep learning-driven incentives can effectively manage the complexities of decentralized edge intelligence. Future work will explore the impact of "malicious actors" (e.g., poisoning attacks), social dynamics between data owners, and dynamic selection of the cluster heads themselves.

Summary updated

Abstract

To enable the large scale and efficient deployment of Artificial Intelligence (AI), the confluence of AI and Edge Computing has given rise to Edge Intelligence, which leverages on the computation and communication capabilities of end devices and edge servers to process data closer to where it is produced. One of the enabling technologies of Edge Intelligence is the privacy preserving machine learning paradigm known as Federated Learning (FL), which enables data owners to conduct model training without having to transmit their raw data to third-party servers. However, the FL network is envisioned to involve thousands of heterogeneous distributed devices. As a result, communication inefficiency remains a key bottleneck. To reduce node failures and device dropouts, the Hierarchical Federated Learning (HFL) framework has been proposed whereby cluster heads are designated to support the data owners through intermediate model aggregation. This decentralized learning approach reduces the reliance on a central controller, e.g., the model owner. However, the issues of resource allocation and incentive design are not well-studied in the HFL framework. In this article, we consider a two-level resource allocation and incentive mechanism design problem. In the lower level, the cluster heads offer rewards in exchange for the data owners' participation, and the data owners are free to choose which cluster to join. Specifically, we apply the evolutionary game theory to model the dynamics of the cluster selection process. In the upper level, each cluster head can choose to serve a model owner, whereas the model owners have to compete amongst each other for the services of the cluster heads. As such, we propose a deep learning based auction mechanism to derive the valuation of each cluster head's services. The performance evaluation shows the uniqueness and stability of our proposed evolutionary game, as well as the revenue maximizing properties of the deep learning based auction.
BibTeX
@article{lim2022decentralized,
  author = {Wei Yang Bryan Lim and Jer Shyuan Ng and Zehui Xiong and Jiangming Jin and Yang Zhang and Dusit Niyato and Cyril Leung and Chunyan Miao},
  title = {{Decentralized Edge Intelligence: A Dynamic Resource Allocation Framework for Hierarchical Federated Learning}},
  journal = {IEEE Trans. Parallel Distributed Syst.},
  year = {2022},
  doi = {10.1109/tpds.2021.3096076},
}
Figure