Multi-Objective Evolutionary Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Federated learning (FL) enables distributed model training without centralizing user data, enhancing privacy.
• Traditional FL models face high communication costs due to frequent global model updates.
• Existing FL methods often struggle to balance communication efficiency and model performance.

Purpose of the Study:

• To optimize neural network structures for federated learning environments.
• To simultaneously minimize communication overhead and global model test errors in FL.
• To enhance the efficiency of evolving deep neural networks within a federated setting.

Main Methods:

• Utilized a multi-objective evolutionary algorithm to optimize neural network architectures.
• Adapted a scalable method for encoding network connectivity tailored for FL.
• Evaluated optimization techniques on multilayer perceptrons and convolutional neural networks.

Main Results:

• The proposed optimization method successfully identified neural network models with reduced communication costs.
• Optimized models demonstrated improved learning performance and accuracy in federated settings.
• Significant reductions in communication overhead were observed compared to standard fully connected networks.

Conclusions:

• Neural network structure optimization is effective in improving federated learning efficiency.
• The multi-objective evolutionary approach balances communication costs and model performance.
• This method offers a promising direction for scalable and efficient federated learning applications.