Assessing Differentially Private Variational Autoencoders Under Membership Inference

Daniel Bernau, Jonas Robl, Florian Kerschbaum

Introduction

Generative models, particularly Variational Autoencoders (VAEs), have become indispensable tools for creating synthetic datasets. However, their tendency to "memorize" training samples poses a significant privacy risk. This memorization allows adversaries to perform Membership Inference (MI) attacks—determining whether a specific individual's data was used to train the model. While Differential Privacy (DP) offers a mathematical framework to mitigate these risks, the practical trade-off between privacy guarantees (the $\epsilon$ parameter) and the utility of the generated data remains difficult to quantify for practitioners.

This paper introduces a standardized framework to evaluate this trade-off. By using MI attacks as a concrete measure of privacy loss, the authors compare various DP mechanisms and their impact on the downstream utility of synthetic data.

Evaluation Framework

The authors propose a three-pillar evaluation strategy to quantify the relationship between privacy and performance:

  1. The Target Model: VAEs are trained using different DP mechanisms, including Central DP (CDP) and Local DP (LDP).
  2. Utility Measurement: Rather than relying solely on reconstruction error, utility is measured by training a "Target Classifier" on synthetic data and evaluating its performance on real, held-out data. This reflects the practical value of the synthetic data for downstream machine learning tasks.
  3. Privacy Measurement: The study employs a "Reconstruction MI Attack." This attack operates on the premise that a VAE will reconstruct training data with significantly lower error than unseen data. If the error is below a certain threshold, the point is inferred to be a member of the training set.

Differential Privacy Strategies

The research compares three distinct approaches to privacy:

  • Central DP (CDP): Implemented via the DP-Adam optimizer, this method adds noise to gradients during the training process to ensure the model weights do not overfit to specific individuals.
  • Local DP (LDP) - Pixelization: A baseline approach where image pixels are perturbed using a Laplace mechanism before the data is used for training.
  • VAE-LDP: A specialized local mechanism that constrains the encoder's output and introduces noise into the latent distribution’s standard deviation during the training phase.

Experimental Setup and Datasets

The framework was tested on two contrasting data types to observe how structural complexity affects privacy:

  • Labeled Faces in the Wild (LFW): High-dimensional image data used for facial recognition (specifically subsets LFW20 and LFW50).
  • MotionSense (MS): A time-series sensor dataset (accelerometer and gyroscope) used for human activity recognition.

Key Insights

Domain-Specific Vulnerability

A primary finding is that image data is significantly "leakier" than time-series data. In baseline scenarios without DP, the reconstruction MI attack achieved perfect accuracy (Average Precision of 1.0) on LFW images. In contrast, the attack struggled with MotionSense data, yielding an AP of only 0.52—barely better than random guessing. This suggests that the high-dimensional structural dependencies in images are more prone to memorization than the noisy, ambiguous patterns in sensor data.

The Regularization Effect of CDP

Central DP acts as a form of regularization. While it inevitably reduces the fidelity of generated samples, it narrows the gap between the model's treatment of training and test data. By "blurring" the specific features an attacker might exploit, CDP provides a direct defense against MI, though often at a high cost to utility.

VAE-LDP: A Potential Sweet Spot

Counter-intuitively, the VAE-LDP mechanism occasionally improved classification accuracy while simultaneously reducing privacy risk. By forcing the model to focus on essential features and ignore peripheral noise, VAE-LDP can produce "cleaner" synthetic datasets. For image tasks, VAE-LDP often outperformed CDP in the relative trade-off metric ($\phi$), maintaining higher utility while successfully pushing the MI attack success rate toward random guessing levels.

Results and Trade-offs

The cost of privacy is most evident in the image domain. For LFW20, applying CDP reduced the MI attack's AP from 1.0 to approximately 0.6. However, this came at a steep expense: the target classifier’s accuracy plummeted from 0.78 to 0.31. The study also noted that for images, the best trade-offs occur at very low noise levels; excessive DP noise tends to destroy structural integrity long before providing additional meaningful privacy gains beyond what a small amount of noise already offers.

Limitations and Future Directions

The effectiveness of VAE-LDP is highly sensitive to hyperparameter tuning, including noise bounds and activation functions. Furthermore, the utility of a generative model is task-dependent; a model optimized for a classifier may not produce visually appealing images for human consumption.

Future research should explore how different VAE architectures, such as $\beta$-VAE, might naturally resist MI attacks through their inherent latent space constraints. Additionally, the failure of standard reconstruction attacks on time-series data highlights the need for more sophisticated, domain-specific MI techniques for non-image data.

Summary updated

BibTeX
@inproceedings{bernau2022assessing,
  author = {Daniel Bernau and Jonas Robl and Florian Kerschbaum},
  title = {{Assessing Differentially Private Variational Autoencoders Under Membership Inference}},
  booktitle = {DBSec},
  year = {2022},
  doi = {10.1007/978-3-031-10684-2\_1},
}
Figure