Enhancing sports image data classification in federated learning through genetic algorithm-based optimization of base...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Federated learning (FL) enables decentralized model training without sharing raw data.
• Traditional FL averages client model weights, potentially limiting optimization.
• Resource-constrained devices pose challenges for deploying complex deep learning models.

Purpose of the Study:

• To introduce a novel method for optimizing base models in federated learning using genetic algorithms (GAs).
• To improve inference time and reduce storage requirements for FL models.
• To enable efficient deployment of FL models on resource-limited hardware.

Main Methods:

• A genetic algorithm approach was developed for optimizing base models in a federated learning setting.
• Key GA operations like chromosome representation, crossover, and mutation were detailed with examples.
• The method was tested using sports datasets across balanced and unbalanced scenarios with varying client numbers.
• Four deep learning architectures (AlexNet, VGG19, ResNet50, EfficientNetB3) were used as base models.

Main Results:

• The GA-based approach significantly improved inference time and reduced storage space.
• Achieved 92.34% accuracy using EfficientNetB3 with 9 clients on a balanced dataset.
• Optimization of EfficientNetB3 resulted in a 20% improvement in inference time and 2.35% reduction in storage space.

Conclusions:

• Genetic algorithms offer an effective strategy for optimizing federated learning models.
• The proposed method enhances model efficiency, making them suitable for resource-constrained environments.
• This research demonstrates the practical benefits of integrating GAs into federated deep learning architectures.