Abstract
Membership inference attacks are used as a key tool for disclosure auditing. They aim to infer whether an individual record was used to train a model. While such evaluations are useful to demonstrate risk, they are computationally expensive and often make strong assumptions about potential adversaries' access to models and training environments, and thus do not provide tight bounds on leakage from potential attacks. We show how prior claims around black-box access being sufficient for optimal membership inference do not hold for stochastic gradient descent, and that optimal membership inference indeed requires white-box access. Our theoretical results lead to a new white-box inference attack, IHA (Inverse Hessian Attack), that explicitly uses model parameters by taking advantage of computing inverse-Hessian vector products. Our results show that both auditors and adversaries may be able to benefit from access to model parameters, and we advocate for further research into white-box methods for membership inference.
Introduction
This research challenges the prevailing "folklore" in the privacy community that black-box access—observing only model outputs like loss or confidence scores—is sufficient for optimal membership inference. By analyzing the discrete-time dynamics of Stochastic Gradient Descent (SGD), we demonstrate that internal model parameters contain unique "fingerprints" of training data that loss values alone cannot reveal.
Historically, the focus on black-box attacks was driven by theoretical results suggesting their optimality. However, those results relied on assumptions (such as parameters following a Boltzmann distribution) that do not hold for models trained with standard SGD. This gap suggests that current privacy audits may significantly underestimate the risk posed by white-box adversaries who have access to model weights.
Taxonomy of Membership Inference Attacks
To situate our work, we compare various membership inference attacks based on their access requirements and the knowledge they utilize.
As shown in the table above, most state-of-the-art attacks like LiRA operate in a black-box setting but require training dozens or hundreds of "shadow" or reference models to establish a baseline. Our proposed method, the Inverse Hessian Attack (IHA), utilizes white-box access to achieve high performance without the massive computational burden of training reference models.
Methodology: The Inverse Hessian Attack (IHA)
We derive a new formulation for the optimal membership inference score by modeling the stationary distribution of SGD. The resulting Inverse Hessian Attack (IHA) examines the influence a specific record has on the model's parameters.
IHA utilizes three primary components:
- Gradients: To identify the direction a record "pulls" the model parameters.
- The Hessian Matrix: To understand the local curvature of the loss landscape.
- Inverse Hessian Vector Products (iHVPs): To calculate how parameters would shift if a specific record were added or removed.
The attack score is composed of four terms ($I_1$ through $I_4$) characterizing a "Newtonian step" for a given record. To ensure scalability for larger models, we utilize Conjugate Gradients (CG) to approximate the inverse Hessian, avoiding the impossible task of computing a full Hessian for high-dimensional parameter spaces.
Experimental Results
We evaluated IHA against standard benchmarks including Purchase-100, MNIST-Odd, and FashionMNIST.
- Superior Performance: On the Purchase-100 tabular dataset, IHA achieved an AUC of 0.703, significantly outperforming the state-of-the-art LiRA attack (0.644).
- High-Confidence Inference: At a strict False Positive Rate (FPR) of 0.1%, IHA achieved a True Positive Rate (TPR) of 7.52%, compared to just 0.98% for LiRA. This indicates that IHA is far more effective at identifying members with high certainty.
- Efficiency: IHA is significantly faster than reference-model-based attacks. For 1,000 samples, IHA required approximately 1 hour of computation, whereas LiRA required 3 hours due to the need to train multiple shadow models.
- Ablation Insights: Our studies found that the $I_2$ term—which measures the alignment between a record's influence and the average influence of the rest of the data—is the primary driver of the attack's success.
Discussion and Limitations
Our findings prove that for SGD-trained models, white-box access provides a distinct advantage for membership inference. Membership is essentially a measure of influence; if a point was in the training set, the model's weights are "aligned" with that point's gradient in a detectable way via the Hessian.
However, challenges remain for scaling these methods. Calculating iHVPs for models with millions or billions of parameters (such as Large Language Models) is computationally intensive. Additionally, the attack's performance is sensitive to the "damping factor" used for Hessian inversion stability. Future work should focus on more efficient iHVP approximations and extending IHA to multi-record inference scenarios.
Summary updated