Introduction
As deep learning models grow in size, concerns regarding data privacy have intensified. Overparameterized models are often viewed as high-capacity "memorizers" that inherently leak information about their training data. This paper challenges that conventional wisdom, demonstrating a "blessing of dimensionality": with appropriate regularization, larger models can simultaneously achieve superior accuracy and enhanced privacy compared to smaller models. By providing a new theoretical framework and extensive empirical evidence, the authors show that the perceived privacy-utility trade-off is not an inescapable law of overparameterization.
The Membership Inference Problem
Membership Inference (MI) attacks aim to determine whether a specific data point was part of a model's training set. These attacks typically exploit the "loss gap"—the phenomenon where a model exhibits significantly lower loss on training samples than on unseen test samples. In the context of overparameterization, the fear is that increased model capacity leads to deeper memorization, making this loss gap wider and the model more vulnerable to MI attacks.
Methodology
The research investigates this phenomenon through two complementary approaches:
- Theoretical Framework: The authors analyze high-dimensional logistic regression using ridge regularization. They employ a novel leave-one-out analysis based on the convex Gaussian min-max theorem (CGMT) to precisely characterize the distribution of model outputs for training versus test points.
- Empirical Validation: The study evaluates deep neural networks across three benchmarks: Purchase100 (classification), CIFAR10 (image recognition), and Multi30K (machine translation). Early stopping is utilized as the primary regularization method, and privacy is measured against a strong "sample-specific loss threshold attack."
The Dynamics of Memorization
A critical insight of the paper is how the distribution of losses evolves during training. As training progresses, the model begins to fit the training data more closely, causing the training loss distribution to shift toward zero while the non-training distribution remains relatively stable. This divergence creates the signal that MI attacks exploit.
As shown in the figure above, the separation between the blue (training) and orange (non-training) histograms becomes more pronounced with more epochs, illustrating why regularization (like early stopping) is vital for privacy.
Key Findings: The Blessing of Dimensionality
The core contribution of the paper is the discovery that wider models have a superior regularization-wise trade-off. While increasing width without adjusting regularization might increase vulnerability, wider models learn generalizable features more efficiently than they memorize specific noise.
Breaking the Privacy-Utility Trade-off
By jointly optimizing model size and regularization, researchers can reach a regime where a larger model is both more accurate and more private. When the number of training epochs is tuned to maintain a specific privacy level, increasing the model width consistently leads to lower generalization error.
As seen in the results across Purchase100, CIFAR10, and Multi30K, increasing the width parameter ($w$) allows the test error to drop significantly. While the raw MI advantage may rise initially, the Pareto frontier for wider models is strictly better: for any target accuracy, a wider model can be made more private than a narrow one through proper regularization.
Conclusion
This work reframes overparameterization from a privacy liability into a potential asset. The "blessing of dimensionality" suggests that the path to more private machine learning does not necessarily require sacrificing performance or using smaller models. Instead, by leveraging the superior learning trajectories of large models and applying rigorous regularization, practitioners can achieve the best of both worlds: high utility and robust data protection.
Summary updated