Benchmarking federated strategies in Peer-to-Peer Federated learning for biomedical data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Growing demand for data protection and privacy fuels research in distributed AI.
• Federated learning (FL) enables collaborative model training on decentralized private data.
• Traditional FL employs a centralized architecture with federated averaging.

Purpose of the Study:

• To evaluate diverse federated learning aggregation strategies in a peer-to-peer environment.
• To identify robust aggregation methods for federated learning, considering varying data sizes.
• To compare proposed strategies against classical federated averaging using biomedical datasets.

Main Methods:

• Implementation of various peer-to-peer federated learning aggregation strategies.
• Development of weighted averaging techniques based on participant contribution and data factors.
• Experimental testing across multiple biomedical datasets with diverse data volumes.

Main Results:

• Accuracy-based weighted averaging demonstrated superior performance compared to federated averaging.
• The proposed strategies showed robustness across different data sizes.
• Federated learning models achieved higher accuracy with optimized aggregation methods.

Conclusions:

• Accuracy-based weighted averaging is a more effective aggregation strategy in peer-to-peer federated learning.
• Peer-to-peer federated learning offers a viable alternative to centralized architectures for privacy-preserving AI.
• Optimized aggregation methods are crucial for enhancing federated learning performance in sensitive data applications.