Privacy-preserving federated neural network learning for disease-associated cell classification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Healthcare Informatics

Background:

• Machine learning model training necessitates extensive data, often siloed in healthcare institutions.
• Privacy regulations impede data centralization or sharing, hindering collaborative research.
• Existing methods struggle to balance data privacy with the need for large, diverse datasets.

Purpose of the Study:

• To introduce PriCell, a privacy-preserving federated learning approach for training complex machine learning models.
• To enable collaborative training of neural networks across multiple healthcare institutions without compromising data confidentiality.
• To ensure patient privacy and data utility for multi-center healthcare studies.

Main Methods:

• Utilized a federated learning framework combined with multiparty homomorphic encryption.
• Developed PriCell for collaborative training of encrypted neural networks.
• Replicated a state-of-the-art convolutional neural network architecture in a decentralized manner.

Main Results:

• Achieved comparable accuracy to centralized, non-secure training methods.
• Successfully preserved the confidentiality of input data, intermediate values, and model parameters.
• Demonstrated efficient, decentralized, and privacy-preserving model training.

Conclusions:

• PriCell offers a robust solution for privacy-preserving federated learning in healthcare.
• The approach guarantees patient privacy while enabling efficient multi-center studies.
• PriCell facilitates the use of complex models like convolutional neural networks on sensitive healthcare data.