A Memory-Efficient Federated Kernel Support Vector Machine for Edge Devices.

Area of Science:

• Machine Learning
• Distributed Systems
• Data Science

Background:

• Federated learning (FL) enables collaborative model training across decentralized devices.
• Kernel Support Vector Machines (SVMs) are powerful classification models but can be memory-intensive.
• Edge devices have limited computational and memory resources, posing challenges for complex model training.

Purpose of the Study:

• To design a federated learning scheme (Fed-KSVM) for training kernel SVMs on edge devices with reduced memory footprint.
• To optimize the training process for kernel SVMs within the constraints of edge computing environments.
• To minimize communication overhead between edge devices and the central server during federated training.

Main Methods:

• Fed-KSVM decomposes kernel SVM training into local training on edge devices using random feature vectors.
• Local optimization problems are divided into subproblems, optimizing subsets of parameters over low-dimensional blocks.
• An incremental learning algorithm, block boosting, solves subproblems sequentially to maintain optimality.
• A global SVM model is constructed by averaging parameters from locally trained models on a central server.

Main Results:

• Fed-KSVM reduces memory consumption on edge devices by approximately 90%.
• The scheme achieves a linear convergence rate, significantly reducing communication costs by up to 99% compared to centralized training.
• Fed-KSVM attained the highest test accuracy among compared state-of-the-art schemes.

Conclusions:

• Fed-KSVM offers an efficient solution for training kernel SVMs in federated learning settings on resource-constrained edge devices.
• The proposed block boosting algorithm effectively reduces memory and communication overhead without sacrificing model accuracy.
• This approach demonstrates the feasibility and effectiveness of advanced federated learning techniques for edge AI applications.