FedPSO: Federated Learning Using Particle Swarm Optimization to Reduce Communication Costs.

Area of Science:

• Machine Learning
• Distributed Systems
• Network Communications

Background:

• Federated learning ensures data privacy by aggregating models on a central server.
• Clients in federated learning often face limited bandwidth and unstable network conditions.
• Existing aggregation methods like FedAvg suffer accuracy degradation due to large weight transmissions in unstable environments.

Purpose of the Study:

• To propose a novel federated learning algorithm, Federated Particle Swarm Optimization (FedPSO).
• To enhance the robustness and communication efficiency of federated learning in unstable network environments.
• To improve the accuracy and reduce communication overhead compared to traditional methods.

Main Methods:

• Replaced the standard FedAvg aggregation with particle swarm optimization.
• Developed FedPSO to transmit score values instead of large model weights between clients and servers.
• Evaluated FedPSO performance in simulated unstable network conditions.

Main Results:

• FedPSO significantly reduced the amount of data transmitted over the network.
• The proposed FedPSO algorithm improved global model accuracy by an average of 9.47%.
• FedPSO demonstrated an approximate 4% improvement in accuracy loss under unstable network conditions.

Conclusions:

• FedPSO offers a robust alternative to FedAvg for federated learning in challenging network environments.
• Transmitting scores instead of weights effectively mitigates accuracy loss due to network instability.
• FedPSO enhances both communication efficiency and model performance in federated learning.