FedMint: Intelligent Bilateral Client Selection in Federated Learning with Newcomer IoT Devices

Osama Wehbi, Sarhad Arisdakessian, Omar Abdel Wahab, Hadi Otrok, Safa Otoum, Azzam Mourad, Mohsen Guizani

Introduction

Federated Learning (FL) has emerged as a cornerstone for privacy-preserving machine learning in the Internet of Things (IoT) era. By allowing models to be trained locally on devices, FL avoids the risks associated with centralized data storage. However, conventional FL frameworks often struggle with "unilateral" selection processes where a central server dictates client participation, frequently ignoring device-side resource constraints and incentives. Furthermore, the "cold start" or newcomer problem—where the server lacks historical performance data for new devices—remains a significant barrier to efficient and fair client selection.

To address these issues, researchers have developed FedMint, a framework that leverages game theory and collaborative bootstrapping to create a bilateral selection process that is both intelligent and inclusive of newcomer devices.

The Challenges of IoT Federated Learning

Standard Federated Learning (VanillaFL) typically employs random client selection, which leads to several inefficiencies:

  1. Resource Heterogeneity: IoT devices vary significantly in computational power (CPU/RAM) and data quality. Random selection often includes "stragglers" that delay global model convergence.
  2. Unilateral Incentives: Most models prioritize server needs. Without considering whether the reward justifies the battery and data costs for the IoT device, participation rates may drop.
  3. The Newcomer Problem: Servers have no baseline to judge the potential contribution of a new device. This often results in high-quality newcomers being excluded or low-quality ones wasting training resources.

The FedMint Framework

FedMint introduces a two-phased approach to modernize client selection.

1. Accuracy Bootstrapping for Newcomers

To solve the "cold start" problem, FedMint utilizes a central bootstrapping server that facilitates knowledge sharing among active federated servers. When a new device joins, the system predicts its likely accuracy based on metadata from similar devices previously encountered by other servers.

The FedMint bootstrapping process for newcomer IoT devices.

The bootstrapping server employs a Decision Tree algorithm using Standard Deviation Reduction (SDR) to map device attributes—such as provider, region, and device type—to expected performance metrics. This allows the system to assign an initial "reputation" or accuracy score to a device before it even begins its first training round.

Decision tree model for predicting device accuracy based on provider and region metadata.

2. Bilateral Matching Game

Once accuracies are established, FedMint treats client selection as a matching game between two sets of rational agents:

  • IoT Devices: These agents rank federated servers based on the monetary rewards offered for participation.
  • Federated Servers: These agents rank devices based on their predicted or historical accuracy and resource availability.

A distributed matching algorithm, inspired by the Gale-Shapley stable marriage model, is used to find a stable pairing. This ensures that no device or server can improve its outcome by unilaterally switching partners, fostering a sustainable and fair data marketplace.

Experimental Evaluation

The framework was evaluated using the MNIST dataset distributed in a non-IID (Independent and Identically Distributed) fashion across a simulated network of over 100 IoT devices. To mimic real-world conditions, devices were assigned varying hardware specifications (300-700 MIPS CPU, 400-900 MB RAM) and network latencies.

Key Results

The implementation of FedMint yielded significant improvements over traditional FL approaches:

  • Enhanced Model Performance: FedMint achieved a global model accuracy of approximately 66% in the initial round, outperforming VanillaFL’s 44%. By the 15th round, FedMint maintained a 20% accuracy lead.
  • Increased Device Incentives: By incorporating device preferences into the selection process, the average monetary earnings for participating IoT devices increased by 58%.
  • Accurate Predictions: The bootstrapping mechanism demonstrated high precision, with a Mean Square Error (MSE) of only ~0.01 when predicting newcomer accuracy.

Conclusion and Future Directions

FedMint demonstrates that bilateralism and metadata-driven bootstrapping can significantly enhance the efficiency and fairness of Federated Learning in IoT environments. By treating client selection as a mutual agreement rather than a server-side command, the framework ensures better model quality and higher device engagement. Future research will focus on decentralizing the bootstrapping server to remove potential bottlenecks and applying differential privacy techniques to the shared metadata to further strengthen security.

Summary updated

BibTeX
@article{wehbi2023fedmint,
  author = {Osama Wehbi and Sarhad Arisdakessian and Omar Abdel Wahab and Hadi Otrok and Safa Otoum and Azzam Mourad and Mohsen Guizani},
  title = {{FedMint: Intelligent Bilateral Client Selection in Federated Learning with Newcomer IoT Devices}},
  journal = {IEEE Internet of Things Journal},
  year = {2023},
  url = {https://openreview.net/forum?id=0yBZF4NH1t},
}
Figure