A non-interactive Online Medical Pre-Diagnosis system on encrypted vertically partitioned data.

Area of Science:

• Medical Informatics
• Cryptography
• Machine Learning

Background:

• Medical records are fragmented across institutions, hindering Online Medical Pre-Diagnosis (OMPD) system development.
• Existing federated learning methods for OMPD involve frequent communication and are vulnerable to inference attacks.
• Vertical data fragmentation in healthcare poses challenges for secure model collaboration.

Purpose of the Study:

• To propose a secure and efficient framework for OMPD systems to overcome vertical data fragmentation.
• To resolve the conflict between medical data isolation and the need for collaborative model training.
• To enhance the security and efficiency of OMPD systems while protecting patient data privacy.

Main Methods:

• Introduction of PPNLR, a secure framework combining functional encryption and blinding factors.
• Development of sample-feature dimension encryption and privacy-preserving vectorization training algorithms.
• Decoupling sample computation from model training for single-round communication between hospitals and cloud servers.

Main Results:

• PPNLR demonstrates resistance to semi-honest inference and collusion attacks.
• Evaluation on six real-world medical datasets shows inference accuracy comparable to centralized plaintext training.
• Achieved at least 3.6x higher computational efficiency and significantly reduced communication complexity compared to existing methods.

Conclusions:

• PPNLR ensures data protection via cryptographic primitives, maintaining high diagnostic accuracy and model parameter security.
• The single-communication architecture lowers deployment barriers in resource-constrained environments.
• PPNLR provides a practical, privacy-friendly framework for building OMPD systems.