GAN-Based Domain Inference Attack

Yuechun Gu, Keke ChenAAAI 2023 · DOI

TL;DR

This paper introduces a novel "Model Domain Inference" (MDI) attack that allows adversaries to infer the likely domain of a target machine learning model, even without prior knowledge of its application. By leveraging Generative Adversarial Networks (GANs), the attack effectively identifies similar domains, enabling attackers to then use data from these domains to significantly boost the success of model-inversion attacks.

What Problem Does This Paper Address?

Machine learning models, especially those trained on sensitive data, pose privacy risks. While membership inference and model inversion attacks are known threats, they typically assume the attacker has "domain knowledge" – understanding the model's application (e.g., facial recognition). This knowledge is crucial for finding relevant auxiliary data to improve attack effectiveness. However, in many real-world scenarios, this domain knowledge is unavailable to the attacker. This paper addresses this gap by proposing a method for adversaries to infer the model's domain using only the model itself and publicly available datasets, thereby enabling them to find useful auxiliary data for subsequent attacks.

Methods

The core of the paper is the GAN-based Model Domain Inference (MDI) attack. The authors propose training a Generative Adversarial Network (GAN) on a "landmark" dataset (a dataset from a known domain). Then, a copy of this GAN is trained further, but this time, the target model's predictions on the GAN-generated images are used to "distract" the GAN's training process.

The intuition is that if a landmark domain is very similar to the target model's latent domain, the target model's feedback will cause less "distraction" during the GAN's training. Conversely, a dissimilar domain will lead to more distraction. The degree of distraction is measured by comparing the output of the original GAN and the "distracted" GAN. A smaller difference indicates a more similar domain.

The paper also briefly explores a model-inversion-based method as an alternative, where a landmark dataset is used as auxiliary data for a model inversion attack, and the quality of the reconstructed data is used to infer domain similarity. However, they find this method to be less effective than the GAN-based approach.

Datasets

The experiments utilize a variety of well-known public datasets for training and evaluation:

  • MNIST
  • EMNIST
  • Fashion-MNIST
  • Clothing
  • Emotion (AffectNet)
  • CIFAR10
  • CIFAR100

For consistency, colored datasets were converted to grayscale and all images were rescaled to 36x36 pixels. The number of classes was unified to seven to align with the Emotion dataset.

Key Insights and Ideas

  • Models implicitly encode domain information: The paper demonstrates that even a black-box target model contains enough information to infer its likely domain, challenging the assumption that domain knowledge is always external.
  • GANs as a tool for domain similarity: The novel use of GANs to measure the "distraction" caused by a target model offers a powerful new way to quantify domain similarity without direct access to the target model's training data.
  • Domain inference unlocks stronger attacks: By enabling attackers to find relevant auxiliary data, the MDI attack significantly enhances the effectiveness of downstream model-inversion attacks.
  • Trade-off between privacy and utility: Mitigation methods like InstaHide and RMT can protect against MDI attacks but come at a slight cost to model accuracy.

Key Results

  • The GAN-based MDI method consistently outperforms the model-inversion-based method in ranking similar domains.
  • Using top-ranked auxiliary datasets identified by the MDI attack can improve model-inversion attack accuracy by up to 30%.
  • Mitigation methods like InstaHide and RMT significantly reduce the effectiveness of the MDI attack, showing strong protection with only a ~5% reduction in model quality.
  • The quality of the WGAN itself can impact the effectiveness of the MDI attack, with simpler datasets (like MNIST) showing more reliable results than complex ones (like CIFAR).

Limitations and Future Work

  • WGAN quality: The effectiveness of the GAN-based MDI attack can be influenced by the underlying WGAN's ability to generate high-quality data for a given dataset. The authors note that WGANs perform less effectively on datasets like CIFAR10 compared to MNIST.
  • Mitigation vulnerabilities: While InstaHide and RMT offer protection, they are susceptible to other types of attacks (e.g., image clustering and pixel sign removing for InstaHide, or regression attacks for RMT under specific conditions).
  • Future Work: The authors suggest extending the study to other learning problems beyond image classification and exploring the use of other generative methods. They also mention the potential for using Fréchet Inception Distance (FID) for grayscale datasets in future work.

Summary updated

BibTeX
@inproceedings{gu2023gan,
  author = {Yuechun Gu and Keke Chen},
  title = {{GAN-Based Domain Inference Attack}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i12.26663},
}
Figure