We give a procedure for computing group-level $(\epsilon, \delta)$-DP guarantees for DP-SGD, when using Poisson sampling or fixed batch size sampling. Up to discretization errors in the implementation, the DP guarantees computed by this procedure are tight (assuming we release every intermediate iterate).
Privacy amplification exploits randomness in data selection to provide tighter differential privacy (DP) guarantees. This analysis is key to DP-SGD's success in machine learning, but, is not readily applicable to the newer state-of-the-art algorithms. This is because these algorithms, known as DP-FTRL, use the matrix mechanism to add correlated noise instead of independent noise as in DP-SGD. In this paper, we propose "MMCC", the first algorithm to analyze privacy amplification via sampling for any generic matrix mechanism. MMCC is nearly tight in that it approaches a lower bound as $\epsilon\to0$. To analyze correlated outputs in MMCC, we prove that they can be analyzed as if they were independent, by conditioning them on prior outputs. Our "conditional composition theorem" has broad utility: we use it to show that the noise added to binary-tree-DP-FTRL can asymptotically match the noise added to DP-SGD with amplification. Our amplification algorithm also has practical empirical utility: we show it leads to significant improvement in the privacy-utility trade-offs for DP-FTRL algorithms on standard benchmarks.
Differentially private learning algorithms inject noise into the learning process. While the most common private learning algorithm, DP-SGD, adds independent Gaussian noise in each iteration, recent work on matrix factorization mechanisms has shown empirically that introducing correlations in the noise can greatly improve their utility. We characterize the asymptotic learning utility for any choice of the correlation function, giving precise analytical bounds for linear regression and as the solution to a convex program for general convex functions. We show, using these bounds, how correlated noise provably improves upon vanilla DP-SGD as a function of problem parameters such as the effective dimension and condition number. Moreover, our analytical expression for the near-optimal correlation function circumvents the cubic complexity of the semi-definite program used to optimize the noise correlation matrix in previous work. We validate our theory with experiments on private deep learning. Our work matches or outperforms prior work while being efficient both in terms of compute and memory.
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.
Differentially private (stochastic) gradient descent is the workhorse of DP private machine learning in both the convex and non-convex settings. Without privacy constraints, second-order methods, like Newton's method, converge faster than first-order methods like gradient descent. In this work, we investigate the prospect of using the second-order information from the loss function to accelerate DP convex optimization. We first develop a private variant of the regularized cubic Newton method of Nesterov and Polyak, and show that for the class of strongly convex loss functions, our algorithm has quadratic convergence and achieves the optimal excess loss. We then design a practical second-order DP algorithm for the unconstrained logistic regression problem. We theoretically and empirically study the performance of our algorithm. Empirical results show our algorithm consistently achieves the best excess loss compared to other baselines and is 10-40x faster than DP-GD/DP-SGD.
We consider the problem of minimizing a non-convex objective while preserving the privacy of the examples in the training data. Building upon the previous variance-reduced algorithm SpiderBoost, we introduce a new framework that utilizes two different kinds of gradient oracles. The first kind of oracles can estimate the gradient of one point, and the second kind of oracles, less precise and more cost-effective, can estimate the gradient difference between two points. SpiderBoost uses the first kind periodically, once every few steps, while our framework proposes using the first oracle whenever the total drift has become large and relies on the second oracle otherwise. This new framework ensures the gradient estimations remain accurate all the time, resulting in improved rates for finding second-order stationary points. Moreover, we address a more challenging task of finding the global minima of a non-convex objective using the exponential mechanism. Our findings indicate that the regularized exponential mechanism can closely match previous empirical and population risk bounds, without requiring smoothness assumptions for algorithms with polynomial running time. Furthermore, by disregarding running time considerations, we show that the exponential mechanism can achieve a good population risk bound and provide a nearly matching lower bound.
In the privacy-utility tradeoff of a model trained on benchmark language and vision tasks, remarkable improvements have been widely reported with the use of pretraining on publicly available data. This is in part due to the benefits of transfer learning, which is the standard motivation for pretraining in non-private settings. However, the stark contrast in the improvement achieved through pretraining under privacy compared to non-private settings suggests that there may be a deeper, distinct cause driving these gains. To explain this phenomenon, we hypothesize that the non-convex loss landscape of a model training necessitates an optimization algorithm to go through two phases. In the first, the algorithm needs to select a good "basin" in the loss landscape. In the second, the algorithm solves an easy optimization within that basin. The former is a harder problem to solve with private data, while the latter is harder to solve with public data due to a distribution shift or data scarcity. Guided by this intuition, we provide theoretical constructions that provably demonstrate the separation between private training with and without public pretraining. Further, systematic experiments on CIFAR10 and LibriSpeech provide supporting evidence for our hypothesis.
All state-of-the-art (SOTA) differentially private machine learning (DP ML) methods are iterative in nature, and their privacy analyses allow publicly releasing the intermediate training checkpoints. However, DP ML benchmarks, and even practical deployments, typically use only the final training checkpoint to make predictions. In this work, for the first time, we comprehensively explore various methods that aggregate intermediate checkpoints to improve the utility of DP training. Empirically, we demonstrate that checkpoint aggregations provide significant gains in the prediction accuracy over the existing SOTA for CIFAR10 and StackOverflow datasets, and that these gains get magnified in settings with periodically varying training data distributions. For instance, we improve SOTA StackOverflow accuracies to 22.7% (+0.43% absolute) for $\epsilon=8.2$, and 23.84% (+0.43%) for $\epsilon=18.9$. Theoretically, we show that uniform tail averaging of checkpoints improves the empirical risk minimization bound compared to the last checkpoint of DP-SGD. Lastly, we initiate an exploration into estimating the uncertainty that DP noise adds in the predictions of DP ML models. We prove that, under standard assumptions on the loss function, the sample variance from last few checkpoints provides a good approximation of the variance of the final model of a DP run. Empirically, we show that the last few checkpoints can provide a reasonable lower bound for the variance of a converged DP model.
In this paper we revisit the problem of differentially private empirical risk minimization (DP-ERM) and stochastic convex optimization (DP-SCO). We show that a well-studied continuous time algorithm from statistical physics called Langevin diffusion (LD) simultaneously provides optimal privacy/utility tradeoffs for both DP-ERM and DP-SCO under $\epsilon$-DP and $(\epsilon,\delta)$-DP. Using the uniform stability properties of LD, we provide the optimal excess population risk guarantee for $\ell_2$-Lipschitz convex losses under $\epsilon$-DP (even up to $\log n$ factors), thus improving on Asi et al. Along the way we provide various technical tools which can be of independent interest: i) A new R\'enyi divergence bound for LD when run on loss functions over two neighboring data sets, ii) Excess empirical risk bounds for last-iterate LD analogous to that of Shamir and Zhang for noisy stochastic gradient descent (SGD), and iii) A two phase excess risk analysis of LD, where the first phase is when the diffusion has not converged in any reasonable sense to a stationary distribution, and in the second phase when the diffusion has converged to a variant of Gibbs distribution. Our universality results crucially rely on the dynamics of LD. When it has converged to a stationary distribution, we obtain the optimal bounds under $\epsilon$-DP. When it is run only for a very short time $\propto 1/p$, we obtain the optimal bounds under $(\epsilon,\delta)$-DP. Here, $p$ is the dimensionality of the model space. Our work initiates a systematic study of DP continuous time optimization. We believe this may have ramifications in the design of discrete time DP optimization algorithms analogous to that in the non-private setting, where continuous time dynamical viewpoints have helped in designing new algorithms, including the celebrated mirror-descent and Polyak's momentum method.
We revisit the problem of using public data to improve the privacy/utility trade-offs for differentially private (DP) model training. Here, public data refers to auxiliary data sets that have no privacy concerns. We consider public data that is from the same distribution as the private training data. For convex losses, we show that a variant of Mirror Descent provides population risk guarantees which are independent of the dimension of the model ($p$). Specifically, we apply Mirror Descent with the loss generated by the public data as the mirror map, and using DP gradients of the loss generated by the private (sensitive) data. To obtain dimension independence, we require $G_Q^2 \leq p$ public data samples, where $G_Q$ is a measure of the isotropy of the loss function. We further show that our algorithm has a natural ``noise stability'' property: If around the current iterate the public loss satisfies $\alpha_v$-strong convexity in a direction $v$, then using noisy gradients instead of the exact gradients shifts our next iterate in the direction $v$ by an amount proportional to $1/\alpha_v$ (in contrast with DP-SGD, where the shift is isotropic). Analogous results in prior works had to explicitly learn the geometry using the public data in the form of preconditioner matrices. Our method is also applicable to non-convex losses, as it does not rely on convexity assumptions to ensure DP guarantees. We demonstrate the empirical efficacy of our algorithm by showing privacy/utility trade-offs on linear regression, deep learning benchmark datasets (WikiText-2, CIFAR-10, and EMNIST), and in federated learning (StackOverflow). We show that our algorithm not only significantly improves over traditional DP-SGD and DP-FedAvg, which do not have access to public data, but also improves over DP-SGD and DP-FedAvg on models that have been pre-trained with the public data to begin with.