On-Demand-FL: A Dynamic and Efficient Multicriteria Federated Learning Client Deployment Scheme

Mario Chahoud, Hani Sami, Azzam Mourad, Safa Otoum, Hadi Otrok, Jamal Bentahar, Mohsen Guizani

TL;DR

This paper introduces "On-Demand-FL," a novel approach to Federated Learning (FL) that dynamically deploys clients as needed, addressing the challenge of device availability in dynamic environments. By leveraging containerization (Docker) and orchestration (Kubernetes), it efficiently selects and deploys clients, leading to faster model convergence, reduced discarded rounds, and improved data utilization.

What Problem Does This Paper Address?

Federated Learning (FL) allows machine learning models to be trained on decentralized data residing on user devices, preserving privacy. However, a significant challenge is the dynamic nature of these devices; they may be unavailable, have limited resources, or experience connectivity issues, leading to dropped training rounds and slower convergence. Existing client selection methods often assume static environments or a fixed set of available clients, failing to adapt to situations where devices are intermittently available or where new devices can join the learning process. This paper tackles the problem of client availability in highly dynamic environments, aiming to maximize data volume and heterogeneity by enabling flexible, on-demand client deployment.

Methods

The core of the proposed solution, On-Demand-FL, lies in its dynamic client deployment and selection strategy. Key components include:

  • Containerization (Docker): Used to create lightweight, efficient environments for deploying FL services on diverse IoT and mobile devices, making them readily available as volunteer clients.
  • Orchestration (Kubernetes): Manages the deployment, scaling, and networking of these containerized services, ensuring efficient resource utilization and dynamic cluster management.
  • Multi-Objective Optimization: The client deployment problem is formulated as a multi-objective optimization problem, considering factors like resource constraints (CPU, memory, battery), availability, location, data diversity, and learning quality.
  • Genetic Algorithm (GA): A heuristic approach is employed to solve the NP-hard multi-objective optimization problem. The GA iteratively evolves a population of potential client deployment solutions, selecting the fittest ones based on the defined objectives.

The architecture involves three layers: the Server (managing the global model and orchestrator requests), the Orchestrator (managing client deployment and profiling), and the Constrained Device (acting as a volunteer client).

Datasets

  • Mobile Data Challenge (MDC) dataset: This dataset, containing GPS-based movement records of users, was used for training and evaluating the proposed approach. It was utilized to simulate real-world scenarios with dynamic device movements and locations.

Key Insights and Ideas

  • Dynamic Client Deployment is Crucial: The paper strongly emphasizes that in dynamic environments, a flexible, on-demand approach to client selection and deployment is essential for efficient FL.
  • Containerization as an Enabler: Docker and Kubernetes are shown to be powerful tools for bridging the gap between resource-constrained devices and the demands of FL, enabling even less capable devices to participate.
  • Multi-Objective Optimization for Realism: Formulating client selection as a multi-objective problem that considers various practical constraints (resource, availability, data quality) leads to more robust and effective solutions.
  • GA for Complex Optimization: The Genetic Algorithm proves to be a suitable heuristic for finding near-optimal solutions in the complex, multi-objective client deployment problem, balancing competing objectives.
  • Data Heterogeneity and Volume: On-demand deployment can significantly increase the volume and diversity of data available for training, which is particularly beneficial for Non-IID datasets common in FL.

Key Results

  • Reduced Discarded Rounds: The On-Demand-FL approach significantly reduced the number of discarded FL rounds compared to default random selection methods.
  • Faster Convergence: The proposed method achieved model convergence in an average of 6-8 learning rounds, compared to 14-18 rounds for the default approach.
  • Increased Client Availability: By dynamically deploying clients, the system demonstrated a higher number of available clients throughout the learning process, especially in environments with initially few available devices.
  • Improved Accuracy: While not the sole focus, the approach contributed to better accuracy, particularly in the early stages of learning, due to the selection of higher-quality clients.
  • Efficient Resource Utilization: The use of containerization and orchestration allows for efficient use of available devices, making them available for FL without requiring constant dedicated resources.

Limitations and Future Work

  • NP-Hard Problem: The client deployment problem is proven to be NP-Hard, meaning finding a perfectly optimal solution is computationally infeasible for large-scale scenarios. The GA provides a heuristic approximation.
  • Security Considerations: The paper acknowledges that selecting nodes as orchestrators raises security issues, but this aspect is outside the scope of the current work.
  • Future Work: The authors plan to extend their architecture to support multiple FL applications running concurrently and to further optimize client deployment. They also mention the need to address the multi-objective optimization problem of selecting orchestrators in future work.

Summary updated

Abstract

In this article, we increase the availability and integration of devices in the learning process to enhance the convergence of federated learning (FL) models. To address the issue of having all the data in one location, FL, which maintains the ability to learn over decentralized data sets, combines privacy and technology. Until the model converges, the server combines the updated weights obtained from each data set over a number of rounds. The majority of the literature suggested client selection techniques to accelerate convergence and boost accuracy. However, none of the existing proposals have focused on the flexibility to deploy and select clients as needed, wherever and whenever that may be. Due to the extremely dynamic surroundings, some devices are actually not available to serve as clients in FL, which affects the availability of data for learning and the applicability of the existing solution for client selection. In this article, we address the aforementioned limitations by introducing an On-Demand-FL, a client deployment approach for FL, offering more volume and heterogeneity of data in the learning process. We make use of the containerization technology, such as Docker, to build efficient environments using Internet of Things and mobile devices serving as volunteers. Furthermore, Kubernetes is used for orchestration. A multiobjective optimization problem representing the client and model deployment is solved using the genetic algorithm (GA) due to its evolutionary strategy. The performed experiments using the mobile data challenge (MDC) data set and the Localfed framework illustrate the relevance of the proposed approach and the efficiency of the on-the-fly deployment of clients whenever and wherever needed with less discarded rounds and more available data.
BibTeX
@article{chahoud2023demand,
  author = {Mario Chahoud and Hani Sami and Azzam Mourad and Safa Otoum and Hadi Otrok and Jamal Bentahar and Mohsen Guizani},
  title = {{On-Demand-FL: A Dynamic and Efficient Multicriteria Federated Learning Client Deployment Scheme}},
  journal = {IEEE Internet of Things Journal},
  year = {2023},
  url = {https://openreview.net/forum?id=HJIiDDgX8B},
}
Figure