Privacy-preserving Generative Framework Against Membership Inference Attacks

Ruikang Yang, Jianfeng Ma, Yinbin Miao, Xindi MaarXiv 2022 · OpenReview

TL;DR

This paper introduces a novel privacy-preserving generative framework (PPGF) that uses a Variational Autoencoder (VAE) to generate synthetic data. This synthetic data, when used to train machine learning models, effectively defends against membership inference attacks while maintaining high utility, offering a new approach to data privacy in ML.

What Problem Does This Paper Address?

The increasing integration of AI and ML into our lives has brought significant privacy concerns. A major vulnerability is the risk of "membership inference attacks," where an adversary can determine if a specific data point was part of a model's training set. This is particularly concerning when sensitive information, like medical diagnoses, is involved. Existing defenses often involve adding noise to model outputs or altering the training process, which can degrade model utility or be difficult to adapt to different models. This paper tackles the challenge of protecting the original training data from such attacks by generating privacy-preserving synthetic data, rather than directly modifying the target model or its outputs.

Methods

The core of the proposed solution is a privacy-preserving generative framework (PPGF) that leverages a Variational Autoencoder (VAE). The process involves:

  1. Information Extraction: The VAE's encoder maps the original source data into a latent space, creating "latent codes." This latent space is designed to capture the essential characteristics of the data.
  2. Privacy-Preserving Noise Addition: Instead of adding noise directly to the data or model, the authors apply a noise mechanism that satisfies a form of "metric privacy" (an extension of differential privacy) to these latent codes. This is a key differentiator, as it operates in the latent space rather than the raw data domain.
  3. Data Reconstruction: The VAE's decoder then reconstructs synthetic data from these noisy latent codes.
  4. Label Transfer (Optional): For supervised learning tasks, a label-transfer mechanism is employed to assign appropriate labels to the newly generated synthetic data.

This approach aims to generate synthetic data that mimics the distribution of the original data while ensuring that the latent representations are sufficiently obfuscated to prevent membership inference.

Datasets

The paper evaluates its framework on three standard image datasets:

  • MNIST: Handwritten digits.
  • Fashion-MNIST: Zalando's article dataset (clothing items).
  • CelebA: Face attributes dataset.

These datasets are commonly used in machine learning research and are suitable for evaluating image classification tasks.

Key Insights and Ideas

  • Latent Space Privacy: The most significant insight is the application of privacy mechanisms (metric privacy) in the latent space of a VAE, rather than directly on the input data or the final model. This allows for a more nuanced approach to privacy preservation.
  • VAE as a Dual Tool: The VAE is ingeniously used for both extracting meaningful features (encoder) and reconstructing data (decoder), acting as a unified component for both information transformation and privacy enforcement.
  • Decoupled Privacy: By generating synthetic data first, the privacy mechanism is decoupled from the target machine learning model. This means the synthetic data can be used to train any downstream model without needing to reconfigure privacy settings for each new model.
  • Balancing Utility and Privacy: The framework demonstrates that it's possible to achieve strong privacy guarantees (low epsilon values) without a drastic loss in data utility, as evidenced by the maintained prediction accuracy of models trained on synthetic data.

Key Results

The experimental evaluation shows promising results:

  • Effective Defense Against Membership Inference Attacks: The PPGF significantly reduces the accuracy of both black-box and white-box membership inference attacks to near random guessing levels (accuracy close to 0.5). This is demonstrated across the MNIST, Fashion-MNIST, and CelebA datasets.
  • High Data Utility: Machine learning models trained on the synthetic data generated by PPGF maintain high prediction accuracy, comparable to models trained on original data, especially when compared to other privacy-preserving methods like DP-SGD. For instance, on MNIST, with a privacy budget $\epsilon = 0.5$, the accuracy drop compared to training on source data is minimal.
  • Comparison with DP-SGD: While DP-SGD also provides privacy, the PPGF often achieves stronger privacy (lower $\epsilon$) with comparable or better utility, especially when the privacy budget is tight. The paper notes that DP-SGD can sometimes achieve higher accuracy but at the cost of significantly weaker privacy (much larger $\epsilon$).

Limitations and Future Work

  • Generative Model Dependency: The utility of the generated synthetic data is inherently limited by the expressive power of the VAE model used. If the VAE cannot accurately capture certain complex features of the original data, the synthetic data might not fully represent them.
  • Fine-grained Noise Mechanism: The authors suggest exploring more sophisticated, fine-grained noise mechanisms to generate even more realistic synthetic data.
  • Application Programming Interface (API): They propose developing an API that allows for the integration of more complex and effective generative models into their framework, making it more versatile.
  • Privacy Budget Trade-off: While the framework balances privacy and utility, there's still a trade-off. Lowering the privacy budget ($\epsilon$) leads to stronger privacy but can slightly decrease data utility.

Summary updated

Abstract

Artificial intelligence and machine learning have been integrated into all aspects of our lives and the privacy of personal data has attracted more and more attention. Since the generation of the model needs to extract the effective information of the training data, the model has the risk of leaking the privacy of the training data. Membership inference attacks can measure the model leakage of source data to a certain degree. In this paper, we design a privacy-preserving generative framework against membership inference attacks, through the information extraction and data generation capabilities of the generative model variational autoencoder (VAE) to generate synthetic data that meets the needs of differential privacy. Instead of adding noise to the model output or tampering with the training process of the target model, we directly process the original data. We first map the source data to the latent space through the VAE model to get the latent code, then perform noise process satisfying metric privacy on the latent code, and finally use the VAE model to reconstruct the synthetic data. Our experimental evaluation demonstrates that the machine learning model trained with newly generated synthetic data can effectively resist membership inference attacks and still maintain high utility.
BibTeX
@misc{yang2022privacy,
  author = {Ruikang Yang and Jianfeng Ma and Yinbin Miao and Xindi Ma},
  title = {{Privacy-preserving Generative Framework Against Membership Inference Attacks}},
  year = {2022},
  eprint = {2202.05469},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.05469},
}
Figure