mDARTS: Searching ML-Based ECG Classifiers Against Membership Inference Attacks.

Area of Science:

• Machine Learning
• Biomedical Signal Processing
• Cybersecurity

Background:

• Electrocardiogram (ECG) classification requires high accuracy and strong privacy.
• Membership inference attacks (MIA) pose a significant privacy risk to neural network models.
• Existing ECG classifiers often struggle to balance performance with privacy preservation.

Purpose of the Study:

• To develop a novel framework for creating ECG classifier architectures that achieve both high classification performance and robust privacy protection.
• To quantify and mitigate privacy leakage in ECG classification neural networks.
• To enhance resilience against membership inference attacks (MIA) in ECG analysis.

Main Methods:

• Development of a privacy estimator to assess and reduce privacy leakage in neural networks.
• Proposal of mDARTS (searching ML-based ECG classifier against MIA), integrating MIA attack loss into architecture search.
• Utilizing heuristic experiments to refine architecture search parameters for ECG classification.

Main Results:

• Achieved 92.1% ECG classification accuracy with a privacy score of 54.3%, indicating reduced information leakage.
• Enhanced classifier performance and privacy scores by up to 3.0% and 1.0% respectively through heuristic experiments.
• Demonstrated framework adaptability for user-customized architectures balancing performance and privacy.

Conclusions:

• The mDARTS framework offers a pioneering solution for high-performance, privacy-preserving ECG classification.
• The study successfully addresses the limitations of previous approaches by integrating privacy considerations into the architecture search process.
• The developed privacy estimator and mDARTS method provide a pathway for creating more secure and effective AI in healthcare.