Fine-Tuning Large Language Models with User-Level Differential Privacy

We investigate practical and scalable algorithms for training large language models (LLMs) with user-level differential privacy (DP) in order to provably safeguard all the examples contributed by each user. We study two variants of DP-SGD with: (1) example-level sampling (ELS) and per-example gradient clipping, and (2) user-level sampling (ULS) and per-user gradient clipping. We derive a novel user-level DP accountant that allows us to compute provably tight privacy guarantees for ELS. Using this, we show that while ELS can outperform ULS in specific settings, ULS generally yields better results when each user has a diverse collection of examples. We validate our findings through experiments in synthetic mean estimation and LLM fine-tuning tasks under fixed compute budgets. We find that ULS is significantly better in settings where either (1) strong privacy guarantees are required, or (2) the compute budget is large. Notably, our focus on LLM-compatible training algorithms allows us to scale to models with hundreds of millions of parameters and datasets with hundreds of thousands of users.

Scaling up the Banded Matrix Factorization Mechanism for Differentially Private ML

DP-BandMF offers a powerful approach to differentially private machine learning, balancing privacy amplification with noise correlation for optimal noise reduction. However, its scalability has been limited to settings where the number of training iterations is less than $10^4$. In this work, we present techniques that significantly extend DP-BandMF's reach, enabling use in settings with and over $10^6$ training iterations. Our enhanced implementation, coupled with extensive experiments, provides clear guidelines on selecting the optimal number of bands. These insights offer practitioners a deeper understanding of DP-BandMF's performance and how to maximize its utility for privacy-preserving machine learning.

Joint Selection: Adaptively Incorporating Public Information for Private Synthetic Data

Mechanisms for generating differentially private synthetic data based on marginals and graphical models have been successful in a wide range of settings. However, one limitation of these methods is their inability to incorporate public data. Initializing a data generating model by pre-training on public data has shown to improve the quality of synthetic data, but this technique is not applicable when model structure is not determined a priori. We develop the mechanism jam-pgm, which expands the adaptive measurements framework to jointly select between measuring public data and private data. This technique allows for public data to be included in a graphical-model-based mechanism. We show that jam-pgm is able to outperform both publicly assisted and non publicly assisted synthetic data generation mechanisms even when the public data distribution is biased.

(Amplified) Banded Matrix Factorization: A unified approach to private training

Matrix factorization (MF) mechanisms for differential privacy (DP) have substantially improved the state-of-the-art in privacy-utility-computation tradeoffs for ML applications in a variety of scenarios, but in both the centralized and federated settings there remain instances where either MF cannot be easily applied, or other algorithms provide better tradeoffs (typically, as $\epsilon$ becomes small). In this work, we show how MF can subsume prior state-of-the-art algorithms in both federated and centralized training settings, across all privacy budgets. The key technique throughout is the construction of MF mechanisms with banded matrices. For cross-device federated learning (FL), this enables multiple-participations with a relaxed device participation schema compatible with practical FL infrastructure (as demonstrated by a production deployment). In the centralized setting, we prove that banded matrices enjoy the same privacy amplification results as for the ubiquitous DP-SGD algorithm, but can provide strictly better performance in most scenarios -- this lets us always at least match DP-SGD, and often outperform it even at $\epsilon\ll2$. Finally, $\hat{b}$-banded matrices substantially reduce the memory and time complexity of per-step noise generation from $\mathcal{O}(n)$, $n$ the total number of iterations, to a constant $\mathcal{O}(\hat{b})$, compared to general MF mechanisms.

Convergence of Gradient Descent with Linearly Correlated Noise and Applications to Differentially Private Learning

We study stochastic optimization with linearly correlated noise. Our study is motivated by recent methods for optimization with differential privacy (DP), such as DP-FTRL, which inject noise via matrix factorization mechanisms. We propose an optimization problem that distils key facets of these DP methods and that involves perturbing gradients by linearly correlated noise. We derive improved convergence rates for gradient descent in this framework for convex and non-convex loss functions. Our theoretical analysis is novel and might be of independent interest. We use these convergence rates to develop new, effective matrix factorizations for differentially private optimization, and highlight the benefits of these factorizations theoretically and empirically.

Relaxed Marginal Consistency for Differentially Private Query Answering

Many differentially private algorithms for answering database queries involve a step that reconstructs a discrete data distribution from noisy measurements. This provides consistent query answers and reduces error, but often requires space that grows exponentially with dimension. Private-PGM is a recent approach that uses graphical models to represent the data distribution, with complexity proportional to that of exact marginal inference in a graphical model with structure determined by the co-occurrence of variables in the noisy measurements. Private-PGM is highly scalable for sparse measurements, but may fail to run in high dimensions with dense measurements. We overcome the main scalability limitation of Private-PGM through a principled approach that relaxes consistency constraints in the estimation objective. Our new approach works with many existing private query answering algorithms and improves scalability or accuracy with no privacy cost.

Graphical-model based estimation and inference for differential privacy

Many privacy mechanisms reveal high-level information about a data distribution through noisy measurements. It is common to use this information to estimate the answers to new queries. In this work, we provide an approach to solve this estimation problem efficiently using graphical models, which is particularly effective when the distribution is high-dimensional but the measurements are over low-dimensional marginals. We show that our approach is far more efficient than existing estimation techniques from the privacy literature and that it can improve the accuracy and scalability of many state-of-the-art mechanisms.

Differentially Private Learning of Undirected Graphical Models using Collective Graphical Models

We investigate the problem of learning discrete, undirected graphical models in a differentially private way. We show that the approach of releasing noisy sufficient statistics using the Laplace mechanism achieves a good trade-off between privacy, utility, and practicality. A naive learning algorithm that uses the noisy sufficient statistics "as is" outperforms general-purpose differentially private learning algorithms. However, it has three limitations: it ignores knowledge about the data generating process, rests on uncertain theoretical foundations, and exhibits certain pathologies. We develop a more principled approach that applies the formalism of collective graphical models to perform inference over the true sufficient statistics within an expectation-maximization framework. We show that this learns better models than competing approaches on both synthetic data and on real human mobility data used as a case study.