Differential Privacy Protection Against Membership Inference Attack on Machine Learning for Genomic Data

Area of Science:

• Genomics
• Machine Learning
• Privacy-Preserving Technologies

Background:

• Genomic data analysis using machine learning (ML) is advancing rapidly.
• Genome privacy is a significant concern, as trained ML models can be vulnerable to privacy breaches like membership inference attacks (MIA).

Purpose of the Study:

• To investigate the vulnerability of ML models to MIA on genomic data.
• To evaluate the effectiveness of differential privacy (DP) as a defense mechanism against MIA.
• To analyze the trade-off between privacy guarantees and model accuracy under DP.

Main Methods:

• Utilized two common ML models: Lasso and convolutional neural networks (CNN).
• Applied differential privacy (DP) with varying privacy budgets to assess its impact on MIA defense and model accuracy.
• Investigated the role of model sparsity in enhancing MIA vulnerability and DP effectiveness.

Main Results:

• The effectiveness of DP in defending against MIA is inversely related to model accuracy; smaller privacy budgets offer stronger privacy but reduce accuracy, following a log-like curve.
• Model sparsity, beyond preventing overfitting, significantly enhances the combined defense of DP against MIA.
• Both DP and model sparsity are crucial for robust genome privacy in ML.

Conclusions:

• ML models for genomic data analysis require robust privacy-preserving techniques.
• Differential privacy and model sparsity are effective strategies to mitigate membership inference attacks.
• Balancing privacy guarantees and predictive accuracy is essential when implementing DP in genomic ML applications.