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 represent a significant privacy risk in machine learning. An adversary executing an MI attack seeks to determine if a specific data point was included in a model's training set. Traditionally, state-of-the-art attacks, such as the Likelihood Ratio Attack (LiRA), rely on loss values. The core assumption is that a model will exhibit lower loss on training data than on unseen data. However, calculating loss typically requires knowledge of the ground-truth label, which may not always be available to an attacker.
This research explores whether model output confidence can serve as a more intuitive and accessible metric for membership inference. Furthermore, it investigates the feasibility of performing these attacks without requiring the true label of the target data point.
Methodology
The proposed approach builds upon the Likelihood Ratio Attack (LiRA) framework but replaces the loss metric with prediction confidence. The process involves:
- Shadow Modeling: Training 16 "shadow models" on random subsets of data. Half of these models include the target point (IN models), while the other half do not (OUT models).
- Metric Selection: Instead of logit-scaled loss, the model's prediction confidence is measured. Gaussian distributions are then fitted to the confidence scores of the IN and OUT model groups.
- Attack Variants:
- Label-Dependent: Uses the log-scaled confidence probability of the actual class:
confidence = log(y_true(f(x)) + 1e-45).
- Label-Dependent: Uses the log-scaled confidence probability of the actual class:
- Label-Independent (Argmax): Uses the highest probabilistic value in the prediction vector, regardless of the class:
confidence = log(argmax(f(x))). - Inference: A parametric likelihood ratio test determines the probability of the target point belonging to the IN distribution versus the OUT distribution.
Experimental Results
The confidence-based metrics demonstrated performance comparable to, and in some cases exceeding, the traditional loss-based LiRA baseline.
As shown in the results, the label-independent method proved remarkably competitive. In the "Attack (Online, Fixed Variance)" setting, the baseline loss method achieved an AUC of 0.5879, while the proposed log(Confidence Values) method reached 0.6009. The log(Argmax) method, which requires no label knowledge, achieved an AUC of 0.5602, validating the viability of label-independent exploits.
Discussion and Future Work
These findings suggest that models naturally exhibit higher certainty on training data due to overfitting on specific patterns and noise. The success of the argmax method is particularly significant, as it removes the requirement for ground-truth labels, making MI attacks more versatile in real-world scenarios.
A primary limitation of this study was computational constraints; shadow models were trained for only 21 epochs with a small sample size (16 models). Future research will focus on scaling the number of shadow models and training epochs to determine the performance ceiling of confidence-based attacks. Additionally, exploring alternative mathematical transformations beyond logarithmic scaling may further improve the fit of confidence values to Gaussian distributions.
Summary updated