Abstract
Significant success has been realized recently on applying machine learning to real-world applications. There have also been corresponding concerns on the privacy of training data, which relates to data security and confidentiality issues. Differential privacy provides a principled and rigorous privacy guarantee on machine learning models. While it is common to design a model satisfying a required differential-privacy property by injecting noise, it is generally hard to balance the trade-off between privacy and utility. We show that stochastic gradient Markov chain Monte Carlo (SG-MCMC) -- a class of scalable Bayesian posterior sampling algorithms proposed recently -- satisfies strong differential privacy with carefully chosen step sizes. We develop theory on the performance of the proposed differentially-private SG-MCMC method. We conduct experiments to support our analysis and show that a standard SG-MCMC sampler without any modification (under a default setting) can reach state-of-the-art performance in terms of both privacy and utility on Bayesian learning.
Introduction
Training machine learning models on sensitive datasets, such as medical records or financial transactions, poses a significant risk of "memorizing" individual data points. Differential Privacy (DP) has emerged as the gold standard for mitigating these risks by ensuring that the inclusion or exclusion of a single individual in the training set does not significantly alter the model's output. However, achieving DP typically requires injecting noise, which often leads to a substantial degradation in model utility—the "privacy-utility trade-off."
Previous attempts to provide DP guarantees for Bayesian sampling methods, specifically Stochastic Gradient Markov Chain Monte Carlo (SG-MCMC), resulted in theoretical step size bounds that were prohibitively small (e.g., $10^{-6}$). Such tiny learning rates make these algorithms impractical for modern deep learning tasks. This work bridges the gap between theory and practice, demonstrating that SG-MCMC naturally satisfies strong DP guarantees with practical step sizes.
Methodology
The study focuses on two primary SG-MCMC variants: Stochastic Gradient Langevin Dynamics (SGLD) and Stochastic Gradient Hamiltonian Monte Carlo (SGHMC). These methods are inherently Bayesian, sampling from the posterior distribution of model weights rather than seeking a single point estimate.
Privacy Bookkeeping and Noise
To achieve tighter privacy bounds, the authors employ the moments accountant method. Unlike standard composition theorems, the moments accountant tracks the privacy loss across iterations more precisely, allowing for a more efficient use of the privacy budget. A critical insight is that the "sampling noise" inherent in SG-MCMC—originally intended to simulate Brownian motion for posterior exploration—can simultaneously serve as the "privacy noise" required for DP.
Gradient Clipping
To bound the influence of any single data point, gradient clipping is applied. By ensuring the norm of the gradients does not exceed a threshold $L$, the sensitivity of the stochastic gradient is controlled, which is a fundamental requirement for establishing DP guarantees.
Theoretical Breakthrough: Practical Step Sizes
A major contribution of this research is the derivation of significantly improved upper bounds for the step size. Previous theoretical frameworks suggested that privacy required infinitesimally small steps. By contrast, this work proves that standard step sizes used in practice (e.g., $\approx 0.1$) are often already sufficient to provide robust $(\epsilon, \delta)$-differential privacy.
As shown in Figure 2, the proposed bounds for both fixed and decreasing step sizes are approximately 1,000 times larger than those previously established, making DP-SGLD viable for complex, large-scale models.
Experimental Results
The proposed DP-SG-MCMC framework was evaluated on standard benchmarks including the Adult, MNIST, and SVHN datasets.
Utility and Accuracy
On the MNIST handwritten digit dataset, DP-SGLD achieved a remarkable 99.12% accuracy under a strict privacy budget ($\epsilon = 0.1$). This represents a negligible drop from the non-private baseline of 99.23%, demonstrating that high utility can be maintained alongside rigorous privacy.
Comparison with State-of-the-Art
The method was also compared against other DP frameworks, including PATE (Private Aggregation of Teacher Ensembles) and OPS (Objective Perturbation for SGD).
As illustrated in Figure 5, DP-SGLD consistently outperforms PATE on the SVHN dataset, particularly in the high-privacy regime (low $\epsilon$). This suggests that the natural regularization provided by Bayesian sampling is highly effective for maintaining utility under privacy constraints.
Conclusion
This work demonstrates that SG-MCMC algorithms are naturally private, providing a "privacy for free" benefit in many Bayesian learning scenarios. By utilizing the moments accountant and refined theoretical bounds, practitioners can achieve state-of-the-art accuracy while maintaining rigorous differential privacy guarantees. Future research will explore the application of this framework to Generative Adversarial Networks (GANs), where the inherent privacy-induced regularization may help mitigate common issues like mode collapse.
Summary updated