Machine Learning with Membership Privacy using Adversarial Regularization

Milad Nasr, Reza Shokri, Amir Houmansadr

Abstract

Membership inference attacks pose a significant threat to the privacy of individuals whose data is used to train machine learning models. This paper introduces a novel adversarial training framework that treats privacy as a min-max game between a classifier and an internal adversary. By explicitly regularizing the model to produce indistinguishable outputs for training and non-training data, the proposed method achieves robust membership privacy with minimal impact on classification accuracy, effectively bridging the gap between heuristic defenses and formal differential privacy.

Introduction

Machine learning models are known to "leak" information about their training data through their output distributions. In a membership inference attack, an adversary observes a model's predictions to determine if a specific data record was part of the training set. This is particularly concerning in sensitive domains like healthcare or finance.

Current mitigation strategies often fall into two extremes: simple heuristics like L2 regularization, which lack formal guarantees, or Differential Privacy (DP), which often significantly degrades model utility. This research proposes Adversarial Regularization as a robust alternative that optimizes for privacy against the strongest possible attacker while maintaining high accuracy.

Methods: Adversarial Regularization

The core contribution is a framework implemented as a min-max game involving two competing neural networks:

  1. The Classifier ($f$): The primary model trained for a specific task (e.g., image classification).
  2. The Inference Model ($h$): An internal adversary that attempts to distinguish between the classifier's predictions on training data ($D$) and a separate reference dataset ($D'$).
The adversarial training framework depicting the min-max game between the classifier and the inference model.

During training, the inference model aims to maximize its ability to identify training set members. Simultaneously, the classifier is trained to minimize its standard task loss plus a penalty based on the inference model's success. This forces the classifier to learn prediction vectors that are statistically indistinguishable across training and non-training samples.

Datasets and Evaluation

The framework was evaluated on three benchmark datasets:

  • CIFAR100: 100-class image recognition.
  • Purchase100: Predicting shopping categories from binary features.
  • Texas100: Predicting patient types from sensitive medical discharge data.

Key Insights

  • Privacy as Indistinguishability: The method treats membership privacy as a distribution matching problem. If the model's behavior is identical for training and general population data, no attack can succeed.
  • Generalization Bonus: By forcing the model to behave consistently on seen and unseen data, adversarial regularization acts as a powerful generalizer, often closing the "generalization gap" more effectively than standard weight-based regularization.

Key Results

The results demonstrate that high privacy can be achieved with low utility costs. On the Purchase100 dataset, attack accuracy was reduced from 67.6% to 51.6% (near the 50% random guess baseline) while maintaining 76.5% classification accuracy. In contrast, achieving similar privacy with L2 regularization caused accuracy to drop to 32.1%.

Impact of adversarial regularization on classifier loss and attacker gain over 50 training epochs for the Purchase100 dataset.

As shown in the figures, the defense stabilizes the classifier loss and prevents the "overfitting" signatures that attackers typically exploit.

Limitations and Future Work

While effective, the method requires a "reference set" from the same distribution as the training data, which may be difficult to obtain in some scenarios. Additionally, adversarial training increases computational complexity. Future research could investigate white-box attack scenarios and methods to reduce the reliance on external reference data.

Summary updated

BibTeX
@inproceedings{nasr2018machine,
  author = {Milad Nasr and Reza Shokri and Amir Houmansadr},
  title = {{Machine Learning with Membership Privacy using Adversarial Regularization}},
  booktitle = {CCS},
  year = {2018},
  doi = {10.1145/3243734.3243855},
}
Figure