Mitigating Membership Inference in Deep Survival Analyses with Differential Privacy

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Deep neural networks are vital for healthcare predictions.
• Sharing trained models aids research but risks data privacy.
• Membership inference attacks can reveal individual data in training sets.

Purpose of the Study:

• To investigate membership leakage in deep survival models.
• To evaluate differential privacy for defending against inference attacks.
• To assess differential privacy's impact on deep survival analysis performance.

Main Methods:

• Assessed membership leakage in deep survival models.
• Developed differentially private training procedures.
• Quantified privacy risks and performance trade-offs.

Main Results:

• Deep survival models were found to leak membership information.
• Differential privacy significantly reduced membership inference risks.
• Differential privacy introduced limited performance loss and potentially enhanced robustness.

Conclusions:

• Deep survival models pose privacy risks.
• Differentially private training offers effective protection for shared models.
• Privacy-preserving methods are crucial for secure collaborative healthcare AI.