Privacy-preserving AUC computation in distributed machine learning with PHT-meDIC

Area of Science:

• Distributed machine learning
• Computational privacy
• Healthcare informatics

Background:

• Computing Area Under the Curve (AUC) in distributed machine learning is challenging due to data privacy concerns.
• Existing cryptographic methods for AUC computation may sacrifice scalability or accuracy.
• Sensitive test data pooling is often restricted in multi-institutional settings.

Purpose of the Study:

• To present novel privacy-preserving solutions for secure AUC computation in distributed environments.
• To enable accurate AUC calculation across multiple institutions without compromising data confidentiality.
• To integrate and demonstrate these methods within a real-world healthcare platform.

Main Methods:

• Developed an exact global AUC method with linear scalability and tie handling.
• Introduced an approximation method for reduced runtime and maintained accuracy.
• Utilized homomorphic encryption (modified Paillier), symmetric/asymmetric cryptography, and randomized encoding.
• Integrated protocols into the Personal Health Train (PHT)-meDIC platform.

Main Results:

• The exact method computes true AUC without revealing private inputs, handling ties effectively.
• The approximation method offers a balance between computational efficiency and precision.
• Demonstrated correctness and feasibility on real-world and synthetic datasets.
• Publicly released code and data for broader adoption.

Conclusions:

• The proposed methods successfully enable privacy-preserving AUC computation in distributed machine learning.
• These solutions are crucial for secure analytics in sensitive domains like healthcare.
• The integration with PHT-meDIC showcases practical applicability and facilitates further research.