Data Masking with Privacy Guarantees

Anh T. Pham, Shalini Ghosh, Vinod YegneswaranarXiv 2019 · OpenReview

Abstract

We study the problem of data release with privacy, where data is made available with privacy guarantees while keeping the usability of the data as high as possible --- this is important in health-care and other domains with sensitive data. In particular, we propose a method of masking the private data with privacy guarantee while ensuring that a classifier trained on the masked data is similar to the classifier trained on the original data, to maintain usability. We analyze the theoretical risks of the proposed method and the traditional input perturbation method. Results show that the proposed method achieves lower risk compared to the input perturbation, especially when the number of training samples gets large. We illustrate the effectiveness of the proposed method of data masking for privacy-sensitive learning on $12$ benchmark datasets.

Introduction

In domains such as healthcare and finance, the ability to share datasets is critical for advancing research and developing robust machine learning models. However, the presence of sensitive individual information often creates a barrier to data dissemination. Traditional anonymization techniques frequently fail to provide rigorous mathematical guarantees, while standard differential privacy methods, such as input perturbation, often degrade data utility to the point of uselessness.

Input perturbation involves adding random noise directly to the raw data before publication. While this protects individual privacy, it distorts the underlying data manifold, causing models trained on this noisy data to suffer significant accuracy drops. This paper addresses a fundamental challenge: how to publish a dataset that satisfies strict differential privacy requirements while remaining highly effective for downstream classification tasks.

Methodology

The authors introduce a novel "reverse-engineering" approach to data masking. Instead of perturbing the input data directly, the method focuses on perturbing the decision boundary of a classifier and then synthesizing a dataset that reflects this private boundary.

Comparison between the proposed masked data generation and the traditional input perturbation method.

The proposed workflow for Logistic Regression consists of three primary steps:

  1. Base Model Training: A standard logistic regression classifier is trained on the original, sensitive dataset to find the optimal weights $w$.
  2. Model Privatization: Laplace noise is added to these weights to produce $w'$, ensuring the model itself satisfies $\epsilon$-differential privacy. This means an adversary cannot reliably infer the presence of any specific individual's data from the model weights.
  3. Masked Data Generation: A new "masked" dataset is iteratively generated. Using backtracking gradient descent, the algorithm synthesizes fake samples such that the gradient of the log-likelihood of this new dataset, with respect to the noisy weights $w'$, is minimized toward zero.

This approach effectively flips the traditional script. By optimizing the synthetic data to fit a pre-calculated, private decision boundary, the resulting dataset retains high utility for classification even if the individual data points look significantly different from the original samples.

Experimental Evaluation

The effectiveness of the proposed method was evaluated using a combination of synthetic data, the MNIST handwritten digit dataset, and 12 UCI benchmark datasets (including Adult Income, Breast Cancer, and Heart Disease).

Visual Proof on MNIST

On the MNIST dataset, the authors demonstrated the ability to replace sensitive classes with visually distinct ones. For instance, images of the digit '3' could be replaced by distorted versions of the digit '6'. Despite this visual transformation, a classifier trained on the masked data remained capable of accurately identifying real '3's in test sets.

Visual representation of masked MNIST digits where sensitive classes are replaced while preserving classification utility.

Performance and Scaling

Across the 12 UCI datasets, the proposed method consistently outperformed standard input perturbation in terms of downstream classifier accuracy across various privacy budgets ($\epsilon$). A key theoretical and empirical insight is the method's behavior relative to dataset size ($N$). While the utility of input perturbation often stagnates as $N$ increases, the proposed method's utility bound tightens, showing improved accuracy as more training samples are available.

Discussion and Conclusion

The proposed data masking technique offers a powerful alternative to traditional privacy-preserving methods. By targeting the decision boundary rather than the raw inputs, it preserves the essential relationships between features and labels while providing a mathematical guarantee of $\epsilon$-differential privacy.

However, some limitations remain. The current implementation is specifically optimized for Logistic Regression (and potentially SVMs). It is yet to be determined how well this synthetic data generalizes to highly non-linear models like Deep Neural Networks. Additionally, the iterative optimization process for data generation is more computationally intensive than simple noise addition. Future research will focus on extending this reverse-engineering framework to more complex, non-convex models and improving the efficiency of the synthesis process.

Summary updated

Abstract

We study the problem of data release with privacy, where data is made available with privacy guarantees while keeping the usability of the data as high as possible --- this is important in health-care and other domains with sensitive data. In particular, we propose a method of masking the private data with privacy guarantee while ensuring that a classifier trained on the masked data is similar to the classifier trained on the original data, to maintain usability. We analyze the theoretical risks of the proposed method and the traditional input perturbation method. Results show that the proposed method achieves lower risk compared to the input perturbation, especially when the number of training samples gets large. We illustrate the effectiveness of the proposed method of data masking for privacy-sensitive learning on $12$ benchmark datasets.
BibTeX
@misc{pham2019data,
  author = {Anh T. Pham and Shalini Ghosh and Vinod Yegneswaran},
  title = {{Data Masking with Privacy Guarantees}},
  year = {2019},
  eprint = {1901.02185},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1901.02185},
}
Figure