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

Junjie Chen, Hui Wang, Xinghua ShibioRxiv 2020 · DOI

TL;DR

This paper explores the privacy risks of using machine learning models on genomic data, specifically focusing on Membership Inference Attacks (MIA) that can reveal if an individual's data was used to train a model. The authors demonstrate that Differential Privacy (DP) can effectively defend against these attacks, albeit with a trade-off in model accuracy, and that model sparsity can further enhance this defense.

What Problem Does This Paper Address?

Genomic data is incredibly valuable for scientific discovery and personalized medicine, but it's also highly sensitive. As machine learning models become increasingly powerful for analyzing this data, concerns arise about how these models themselves can inadvertently leak private information. One such threat is a Membership Inference Attack (MIA), where an attacker, by querying a trained model, can determine if a specific individual's genomic data was part of the training set. This is particularly concerning for genomic data, as it can reveal sensitive information like disease susceptibility or family history. While Differential Privacy (DP) is a known defense against MIA in other domains, its effectiveness and implications for genomic data, especially when combined with common machine learning techniques like Lasso and CNNs, remain underexplored.

Methods

The researchers investigate two common machine learning models: Lasso (a sparse linear model) and Convolutional Neural Networks (CNNs). They simulate a white-box MIA scenario, meaning the attacker has full knowledge of the target model's architecture and hyperparameters. To defend against these attacks, they implement Differential Privacy (DP) by adding calibrated noise to the gradients during the model training process. The core of their approach involves:

  • MIA Simulation: Building "shadow models" that mimic the target model's behavior on a separate dataset to train an "attack model" that predicts membership.
  • Differential Privacy Implementation: Using TensorFlow-privacy to integrate DP into the training of both Lasso and CNN models. This involves clipping gradients and adding Gaussian noise, controlled by a privacy budget ($\epsilon$).
  • Trade-off Analysis: Quantitatively evaluating the relationship between the privacy budget ($\epsilon$) and the target model's prediction accuracy.
  • Sparsity Investigation: Examining how introducing model sparsity (using L1 regularization in Lasso and CNNs) affects both vulnerability to MIA and the effectiveness of DP as a defense.

Datasets

The study utilizes a publicly available yeast genomic dataset. This dataset contains genotypes for 28,820 genetic variants from 4,390 individuals, along with 20 end-point growth traits. For their experiments, they focus on one specific trait, "Copper Sulfate," and binarize it for classification tasks. The dataset is split into a private target dataset and a public shadow dataset for training and evaluation.

Key Insights and Ideas

  • Genomic Data is Vulnerable to MIA: The study confirms that even without DP, MIA can successfully infer membership in training datasets for models trained on genomic data.
  • DP is a Viable Defense: Differential Privacy effectively mitigates MIA on genomic data, reducing the attack's accuracy.
  • Privacy-Accuracy Trade-off is Inevitable: Implementing DP to enhance privacy inevitably leads to a decrease in the target model's prediction accuracy. This trade-off is crucial for practical deployment.
  • Sparsity Enhances Defense: Introducing model sparsity (e.g., through L1 regularization) not only helps prevent overfitting in high-dimensional genomic data but also significantly boosts the effectiveness of DP in defending against MIA.
  • Log-like Relationship: The relationship between the privacy budget ($\epsilon$) and model accuracy follows a log-like curve, suggesting a critical "sweet spot" for balancing privacy and utility.

Key Results

  • MIA Effectiveness: Without DP, MIA achieved an accuracy of around 57% for both Lasso and CNN models on the yeast genomic data, significantly better than random guessing (50%).
  • DP's Impact on Accuracy: As the privacy budget ($\epsilon$) decreases (meaning stronger privacy), the target model's accuracy deteriorates. For instance, with a small $\epsilon$, accuracy dropped noticeably.
  • DP's Impact on MIA: As the privacy budget ($\epsilon$) decreases, the attack accuracy of MIA also decreases, indicating DP's effectiveness. For Lasso, attack accuracy dropped significantly with $\epsilon < 10$. For CNN, it steadily decreased as $\epsilon$ increased.
  • Sparsity Benefits: Introducing sparsity improved the target model's accuracy (from 0.7910 to 0.7963 for Lasso) and reduced MIA attack accuracy (from 0.5728 to 0.5631 for Lasso) when DP was not deployed. When DP was deployed, sparse models showed better privacy protection for the same DP budget.
  • Optimal Privacy Budget: The study suggests a privacy budget of $\epsilon = 10$ as a good balance between privacy and accuracy for the investigated dataset and models.

Limitations and Future Work

  • White-Box Attack Model: The current study assumes a white-box attack scenario. Future work should explore black-box MIA, where the attacker only has query access to the target model.
  • Yeast Data: The experiments were conducted on yeast genomic data. Extending the framework to large-scale human genomic data is a crucial next step.
  • Limited Model Types: The study focused on Lasso and CNNs. Investigating other machine learning models, including GANs and their variants, is important.
  • Specific Genomic Analysis: The focus was on phenotype prediction. Future work should explore MIA and DP in other genomic analysis tasks like association studies or risk prediction.
  • Hyperparameter Tuning: The study used a specific value for sparsity ($\lambda = 0.001352$). Further exploration of different sparsity levels and their interaction with DP could be beneficial.

Summary updated

Abstract

Machine learning is powerful to model massive genomic data while genome privacy is a growing concern. Studies have shown that not only the raw data but also the trained model can potentially infringe genome privacy. An example is the membership inference attack (MIA), by which the adversary, who only queries a given target model without knowing its internal parameters, can determine whether a specific record was included in the training dataset of the target model. Differential privacy (DP) has been used to defend against MIA with rigorous privacy guarantee. In this paper, we investigate the vulnerability of machine learning against MIA on genomic data, and evaluate the effectiveness of using DP as a defense mechanism. We consider two widely-used machine learning models, namely Lasso and convolutional neural network (CNN), as the target model. We study the trade-off between the defense power against MIA and the prediction accuracy of the target model under various privacy settings of DP. Our results show that the relationship between the privacy budget and target model accuracy can be modeled as a log-like curve, thus a smaller privacy budget provides stronger privacy guarantee with the cost of losing more model accuracy. We also investigate the effect of model sparsity on model vulnerability against MIA. Our results demonstrate that in addition to prevent overfitting, model sparsity can work together with DP to significantly mitigate the risk of MIA.
BibTeX
@misc{chen2020differential,
  author = {Junjie Chen and Hui Wang and Xinghua Shi},
  title = {{Differential Privacy Protection Against Membership Inference Attack on Machine Learning for Genomic Data}},
  year = {2020},
  doi = {10.1101/2020.08.03.235416},
  howpublished = {bioRxiv},
}
Figure