Confidence Is All You Need for MI Attacks

Abhishek Sinha, Himanshi Tibrewal, Mansi Gupta, Nikhar Waghela, Shivank Garg

Abstract

In this evolving era of machine learning security, membership inference attacks have emerged as a potent threat to the confidentiality of sensitive data. In this attack, adversaries aim to determine whether a particular point was used during the training of a target model. This paper proposes a new method to gauge a data point's membership in a model's training set. Instead of correlating loss with membership, as is traditionally done, we have leveraged the fact that training examples generally exhibit higher confidence values when classified into their actual class. During training, the model is essentially being 'fit' to the training data and might face particular difficulties in generalization to unseen data. This asymmetry leads to the model achieving higher confidence on the training data as it exploits the specific patterns and noise present in the training data. Our proposed approach leverages the confidence values generated by the machine learning model. These confidence values provide a probabilistic measure of the model's certainty in its predictions and can further be used to infer the membership of a given data point. Additionally, we also introduce another variant of our method that allows us to carry out this attack without knowing the ground truth (true class) of a given data point, thus offering an edge over existing label-dependent attack methods.

Introduction

Membership Inference (MI) attacks are a major privacy concern in machine learning. An adversary tries to determine if a specific individual's data (e.g., a medical record or a private photo) was used to train a target model. If successful, this leaks sensitive information about the training set.

Most existing attacks, like the Likelihood Ratio Attack (LiRA), rely on loss values. The logic is that models minimize loss on training data, so a low loss suggests membership. However, loss-based attacks often require knowing the "ground truth" label of a data point to calculate the error. This paper explores whether the raw confidence (the probability assigned to a class) is a more direct and versatile signal for membership, especially when labels are unknown.

Methodology

The authors build upon the framework established by Carlini et al. (2022), known as LiRA. Their process involves:

  1. Shadow Models: Training 16 "shadow models" on random subsets of data. For a specific target point $(x, y)$, half the models include it in their training (IN models) and half do not (OUT models).
  2. Confidence Metrics: Instead of calculating loss, they extract the model's confidence. They propose two main metrics:
  3. Label-Dependent: $log(P(y_{true} | x))$, which is the log-probability the model assigns to the actual correct class.
  4. Label-Independent (Argmax): $log(max(P(x)))$, which is the log-probability of the model's most confident prediction, regardless of whether it's correct.
  5. Likelihood Ratio Test: They fit Gaussian distributions to the confidence scores of the IN and OUT models. To test a new point, they compare its confidence score against these distributions to calculate the likelihood of it being a "member."

Key Results

The authors compared their confidence-based methods against a LiRA baseline using the Area Under the Curve (AUC) metric across various attack types (Online, Offline, and Global Threshold).

Visual comparison of AUC results across different attack configurations and confidence metrics.

As shown in Figure 1 and Table 1, the confidence-based metrics remained on par with or slightly exceeded traditional loss-based approaches:

  • Comparable Performance: The "Attack (Online, Fixed Variance)" using log-confidence achieved an AUC of 0.6009, slightly outperforming the loss-based baseline of 0.5879.
  • Label-Independent Success: The argmax approach (which requires no ground truth labels) achieved an AUC of 0.5622, proving that MI attacks remain a viable threat even without label information.
Detailed AUC values for different membership inference attack variants and metrics.

Discussion and Insights

The core insight is that during training, a model is essentially being "fit" to its training data. This causes an asymmetry: the model exploits specific patterns and noise in the training set, leading to higher certainty (confidence) on those points compared to unseen data. By using the argmax as a proxy for confidence, the authors demonstrate that an attacker doesn't need to know the true label of a data point to perform a successful MI attack. This makes the attack much more practical in real-world scenarios where labels might be private or unknown.

Limitations and Future Work

The authors acknowledge that their results were limited by hardware, using only 16 shadow models and training for 21 epochs. Scaling up the number of shadow models and increasing training time would likely significantly improve the attack's accuracy. Future work could explore other mathematical transformations to better model the Gaussian distribution of confidence scores and apply this confidence-based methodology to other types of MI attacks beyond the LiRA framework.

Summary updated

Abstract

In this evolving era of machine learning security, membership inference attacks have emerged as a potent threat to the confidentiality of sensitive data. In this attack, adversaries aim to determine whether a particular point was used during the training of a target model. This paper proposes a new method to gauge a data point's membership in a model's training set. Instead of correlating loss with membership, as is traditionally done, we have leveraged the fact that training examples generally exhibit higher confidence values when classified into their actual class. During training, the model is essentially being 'fit' to the training data and might face particular difficulties in generalization to unseen data. This asymmetry leads to the model achieving higher confidence on the training data as it exploits the specific patterns and noise present in the training data. Our proposed approach leverages the confidence values generated by the machine learning model. These confidence values provide a probabilistic measure of the model's certainty in its predictions and can further be used to infer the membership of a given data point. Additionally, we also introduce another variant of our method that allows us to carry out this attack without knowing the ground truth(true class) of a given data point, thus offering an edge over existing label-dependent attack methods.
BibTeX
@inproceedings{sinha2024confidence,
  author = {Abhishek Sinha and Himanshi Tibrewal and Mansi Gupta and Nikhar Waghela and Shivank Garg},
  title = {{Confidence Is All You Need for MI Attacks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=kednjQ9nkz},
}
Figure