Om Thakkar

Matthew Jagielski, Om Thakkar, Florian Tramèr et al. · berkeley, eth-zurich

Machine learning models exhibit two seemingly contradictory phenomena: training data memorization, and various forms of forgetting. In memorization, models overfit specific training examples and become susceptible to privacy attacks. In forgetting, examples which appeared early in training are forgotten by the end. In this work, we connect these phenomena. We propose a technique to measure to what extent models "forget" the specifics of training examples, becoming less susceptible to privacy attacks on examples they have not seen recently. We show that, while non-convex models can memorize data forever in the worst-case, standard image, speech, and language models empirically do forget examples over time. We identify nondeterminism as a potential explanation, showing that deterministically trained models do not forget. Our results suggest that examples seen early when training with extremely large datasets - for instance those examples used to pre-train a model - may observe privacy benefits at the expense of examples seen later.

Virat Shejwalkar, Arun Ganesh, Rajiv Mathews et al. · cmu

In this work, we focus on improving the accuracy-variance trade-off for state-of-the-art differentially private machine learning (DP ML) methods. First, we design a general framework that uses aggregates of intermediate checkpoints \emph{during training} to increase the accuracy of DP ML techniques. Specifically, we demonstrate that training over aggregates can provide significant gains in prediction accuracy over the existing state-of-the-art for StackOverflow, CIFAR10 and CIFAR100 datasets. For instance, we improve the state-of-the-art DP StackOverflow accuracies to 22.74\% (+2.06\% relative) for $ε=8.2$, and 23.90\% (+2.09\%) for $ε=18.9$. Furthermore, these gains magnify in settings with periodically varying training data distributions. We also demonstrate that our methods achieve relative improvements of 0.54\% and 62.6\% in terms of utility and variance, on a proprietary, production-grade pCVR task. Lastly, we initiate an exploration into estimating the uncertainty (variance) 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. Crucially, all the methods proposed in this paper operate on \emph{a single training run} of the DP ML technique, thus incurring no additional privacy cost.

Arun Ganesh, Mahdi Haghifam, Milad Nasr et al.

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.

Ehsan Amid, Om Thakkar, Arun Narayanan et al.

Recent work has designed methods to demonstrate that model updates in ASR training can leak potentially sensitive attributes of the utterances used in computing the updates. In this work, we design the first method to demonstrate information leakage about training data from trained ASR models. We design Noise Masking, a fill-in-the-blank style method for extracting targeted parts of training data from trained ASR models. We demonstrate the success of Noise Masking by using it in four settings for extracting names from the LibriSpeech dataset used for training a state-of-the-art Conformer model. In particular, we show that we are able to extract the correct names from masked training utterances with 11.8% accuracy, while the model outputs some name from the train set 55.2% of the time. Further, we show that even in a setting that uses synthetic audio and partial transcripts from the test set, our method achieves 2.5% correct name accuracy (47.7% any name success rate). Lastly, we design Word Dropout, a data augmentation method that we show when used in training along with Multistyle TRaining (MTR), provides comparable utility as the baseline, along with significantly mitigating extraction via Noise Masking across the four evaluated settings.

Geeticka Chauhan, Steve Chien, Om Thakkar et al.

Self-supervised learning (SSL) methods for large speech models have proven to be highly effective at ASR. With the interest in public deployment of large pre-trained models, there is a rising concern for unintended memorization and leakage of sensitive data points from the training data. In this paper, we apply differentially private (DP) pre-training to a SOTA Conformer-based encoder, and study its performance on a downstream ASR task assuming the fine-tuning data is public. This paper is the first to apply DP to SSL for ASR, investigating the DP noise tolerance of the BEST-RQ pre-training method. Notably, we introduce a novel variant of model pruning called gradient-based layer freezing that provides strong improvements in privacy-utility-compute trade-offs. Our approach yields a LibriSpeech test-clean/other WER (%) of 3.78/ 8.41 with ($10$, 1e^-9)-DP for extrapolation towards low dataset scales, and 2.81/ 5.89 with (10, 7.9e^-11)-DP for extrapolation towards high scales.

W. Ronny Huang, Steve Chien, Om Thakkar et al.

End-to-end (E2E) models are often being accompanied by language models (LMs) via shallow fusion for boosting their overall quality as well as recognition of rare words. At the same time, several prior works show that LMs are susceptible to unintentionally memorizing rare or unique sequences in the training data. In this work, we design a framework for detecting memorization of random textual sequences (which we call canaries) in the LM training data when one has only black-box (query) access to LM-fused speech recognizer, as opposed to direct access to the LM. On a production-grade Conformer RNN-T E2E model fused with a Transformer LM, we show that detecting memorization of singly-occurring canaries from the LM training data of 300M examples is possible. Motivated to protect privacy, we also show that such memorization gets significantly reduced by per-example gradient-clipped LM training without compromising overall quality.

Lun Wang, Om Thakkar, Rajiv Mathews

It is well-known that neural networks can unintentionally memorize their training examples, causing privacy concerns. However, auditing memorization in large non-auto-regressive automatic speech recognition (ASR) models has been challenging due to the high compute cost of existing methods such as hardness calibration. In this work, we design a simple auditing method to measure memorization in large ASR models without the extra compute overhead. Concretely, we speed up randomly-generated utterances to create a mapping between vocal and text information that is difficult to learn from typical training examples. Hence, accurate predictions only for sped-up training examples can serve as clear evidence for memorization, and the corresponding accuracy can be used to measure memorization. Using the proposed method, we showcase memorization in the state-of-the-art ASR models. To mitigate memorization, we tried gradient clipping during training to bound the influence of any individual example on the final model. We empirically show that clipping each example's gradient can mitigate memorization for sped-up training examples with up to 16 repetitions in the training set. Furthermore, we show that in large-scale distributed training, clipping the average gradient on each compute core maintains neutral model quality and compute cost while providing strong privacy protection.

Erchi Wang, Pengrun Huang, Eli Chien et al.

Differential privacy (DP) has a wide range of applications for protecting data privacy, but designing and verifying DP algorithms requires expert-level reasoning, creating a high barrier for non-expert practitioners. Prior works either rely on specialized verification languages that demand substantial domain expertise or remain semi-automated and require human-in-the-loop guidance. In this work, we investigate whether large language models (LLMs) can automate DP reasoning. We introduce DPrivBench, a benchmark in which each instance asks whether a function or algorithm satisfies a stated DP guarantee under specified assumptions. The benchmark is carefully designed to cover a broad range of DP topics, span diverse difficulty levels, and resist shortcut reasoning through trivial pattern matching. Experiments show that while the strongest models handle textbook mechanisms well, all models struggle with advanced algorithms, revealing substantial gaps in current DP reasoning capabilities. Through further analytic study and failure-mode analysis, we identify several promising directions for improving automated DP reasoning. Our benchmark provides a solid foundation for developing and evaluating such methods, and complements existing benchmarks for mathematical reasoning.

Peter Kairouz, Brendan McMahan, Shuang Song et al.

We consider training models with differential privacy (DP) using mini-batch gradients. The existing state-of-the-art, Differentially Private Stochastic Gradient Descent (DP-SGD), requires privacy amplification by sampling or shuffling to obtain the best privacy/accuracy/computation trade-offs. Unfortunately, the precise requirements on exact sampling and shuffling can be hard to obtain in important practical scenarios, particularly federated learning (FL). We design and analyze a DP variant of Follow-The-Regularized-Leader (DP-FTRL) that compares favorably (both theoretically and empirically) to amplified DP-SGD, while allowing for much more flexible data access patterns. DP-FTRL does not use any form of privacy amplification. The code is available at https://github.com/google-research/federated/tree/master/dp_ftrl and https://github.com/google-research/DP-FTRL .

Matthew Jagielski, Om Thakkar, Lun Wang

Speech models are often trained on sensitive data in order to improve model performance, leading to potential privacy leakage. Our work considers noise masking attacks, introduced by Amid et al. 2022, which attack automatic speech recognition (ASR) models by requesting a transcript of an utterance which is partially replaced with noise. They show that when a record has been seen at training time, the model will transcribe the noisy record with its memorized sensitive transcript. In our work, we extend these attacks beyond ASR models, to attack pretrained speech encoders. Our method fine-tunes the encoder to produce an ASR model, and then performs noise masking on this model, which we find recovers private information from the pretraining data, despite the model never having seen transcripts at pretraining time! We show how to improve the precision of these attacks and investigate a number of countermeasures to our attacks.

Lun Wang, Om Thakkar, Zhong Meng et al.

Gradient clipping plays a vital role in training large-scale automatic speech recognition (ASR) models. It is typically applied to minibatch gradients to prevent gradient explosion, and to the individual sample gradients to mitigate unintended memorization. This work systematically investigates the impact of a specific granularity of gradient clipping, namely per-core clip-ping (PCC), across training a wide range of ASR models. We empirically demonstrate that PCC can effectively mitigate unintended memorization in ASR models. Surprisingly, we find that PCC positively influences ASR performance metrics, leading to improved convergence rates and reduced word error rates. To avoid tuning the additional hyperparameter introduced by PCC, we further propose a novel variant, adaptive per-core clipping (APCC), for streamlined optimization. Our findings highlight the multifaceted benefits of PCC as a strategy for robust, privacy-forward ASR model training.

Ehsan Amid, Arun Ganesh, Rajiv Mathews et al.

In this paper, we revisit the problem of using in-distribution 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 design a natural variant of DP mirror descent, where the DP gradients of the private/sensitive data act as the linear term, and the loss generated by the public data as the mirror map. We show that, for linear regression with feature vectors drawn from a non-isotropic sub-Gaussian distribution, our algorithm, PDA-DPMD (a variant of mirror descent), provides population risk guarantees that are asymptotically better than the best known guarantees under DP (without having access to public data), when the number of public data samples ($n_{\sf pub}$) is sufficiently large. We further show that our algorithm has natural "noise stability" properties that control the variance due to noise added to ensure DP. We demonstrate the efficacy of our algorithm by showing privacy/utility trade-offs on four benchmark datasets (StackOverflow, WikiText-2, CIFAR-10, and EMNIST). We show that our algorithm not only significantly improves over traditional DP-SGD, which does not have access to public data, but to our knowledge is the first to improve over DP-SGD on models that have been pre-trained with public data.

Shubhankar Mohapatra, Sajin Sasy, Xi He et al.

Hyperparameter optimization is a ubiquitous challenge in machine learning, and the performance of a trained model depends crucially upon their effective selection. While a rich set of tools exist for this purpose, there are currently no practical hyperparameter selection methods under the constraint of differential privacy (DP). We study honest hyperparameter selection for differentially private machine learning, in which the process of hyperparameter tuning is accounted for in the overall privacy budget. To this end, we i) show that standard composition tools outperform more advanced techniques in many settings, ii) empirically and theoretically demonstrate an intrinsic connection between the learning rate and clipping norm hyperparameters, iii) show that adaptive optimizers like DPAdam enjoy a significant advantage in the process of honest hyperparameter tuning, and iv) draw upon novel limiting behaviour of Adam in the DP setting to design a new and more efficient optimizer.

Trung Dang, Om Thakkar, Swaroop Ramaswamy et al.

Distributed learning paradigms such as federated learning often involve transmission of model updates, or gradients, over a network, thereby avoiding transmission of private data. However, it is possible for sensitive information about the training data to be revealed from such gradients. Prior works have demonstrated that labels can be revealed analytically from the last layer of certain models (e.g., ResNet), or they can be reconstructed jointly with model inputs by using Gradients Matching [Zhu et al'19] with additional knowledge about the current state of the model. In this work, we propose a method to discover the set of labels of training samples from only the gradient of the last layer and the id to label mapping. Our method is applicable to a wide variety of model architectures across multiple domains. We demonstrate the effectiveness of our method for model training in two domains - image classification, and automatic speech recognition. Furthermore, we show that existing reconstruction techniques improve their efficacy when used in conjunction with our method. Conversely, we demonstrate that gradient quantization and sparsification can significantly reduce the success of the attack.

Trung Dang, Om Thakkar, Swaroop Ramaswamy et al.

End-to-end Automatic Speech Recognition (ASR) models are commonly trained over spoken utterances using optimization methods like Stochastic Gradient Descent (SGD). In distributed settings like Federated Learning, model training requires transmission of gradients over a network. In this work, we design the first method for revealing the identity of the speaker of a training utterance with access only to a gradient. We propose Hessian-Free Gradients Matching, an input reconstruction technique that operates without second derivatives of the loss function (required in prior works), which can be expensive to compute. We show the effectiveness of our method using the DeepSpeech model architecture, demonstrating that it is possible to reveal the speaker's identity with 34% top-1 accuracy (51% top-5 accuracy) on the LibriSpeech dataset. Further, we study the effect of two well-known techniques, Differentially Private SGD and Dropout, on the success of our method. We show that a dropout rate of 0.2 can reduce the speaker identity accuracy to 0% top-1 (0.5% top-5).

Swaroop Ramaswamy, Om Thakkar, Rajiv Mathews et al.

This paper presents the first consumer-scale next-word prediction (NWP) model trained with Federated Learning (FL) while leveraging the Differentially Private Federated Averaging (DP-FedAvg) technique. There has been prior work on building practical FL infrastructure, including work demonstrating the feasibility of training language models on mobile devices using such infrastructure. It has also been shown (in simulations on a public corpus) that it is possible to train NWP models with user-level differential privacy using the DP-FedAvg algorithm. Nevertheless, training production-quality NWP models with DP-FedAvg in a real-world production environment on a heterogeneous fleet of mobile phones requires addressing numerous challenges. For instance, the coordinating central server has to keep track of the devices available at the start of each round and sample devices uniformly at random from them, while ensuring \emph{secrecy of the sample}, etc. Unlike all prior privacy-focused FL work of which we are aware, for the first time we demonstrate the deployment of a differentially private mechanism for the training of a production neural network in FL, as well as the instrumentation of the production training infrastructure to perform an end-to-end empirical measurement of unintended memorization.

Borja Balle, Peter Kairouz, H. Brendan McMahan et al.

Differentially Private Stochastic Gradient Descent (DP-SGD) forms a fundamental building block in many applications for learning over sensitive data. Two standard approaches, privacy amplification by subsampling, and privacy amplification by shuffling, permit adding lower noise in DP-SGD than via naïve schemes. A key assumption in both these approaches is that the elements in the data set can be uniformly sampled, or be uniformly permuted -- constraints that may become prohibitive when the data is processed in a decentralized or distributed fashion. In this paper, we focus on conducting iterative methods like DP-SGD in the setting of federated learning (FL) wherein the data is distributed among many devices (clients). Our main contribution is the \emph{random check-in} distributed protocol, which crucially relies only on randomized participation decisions made locally and independently by each client. It has privacy/accuracy trade-offs similar to privacy amplification by subsampling/shuffling. However, our method does not require server-initiated communication, or even knowledge of the population size. To our knowledge, this is the first privacy amplification tailored for a distributed learning framework, and it may have broader applicability beyond FL. Along the way, we extend privacy amplification by shuffling to incorporate $(ε,δ)$-DP local randomizers, and exponentially improve its guarantees. In practical regimes, this improvement allows for similar privacy and utility using data from an order of magnitude fewer users.

Om Thakkar, Swaroop Ramaswamy, Rajiv Mathews et al.

Recent works have shown that generative sequence models (e.g., language models) have a tendency to memorize rare or unique sequences in the training data. Since useful models are often trained on sensitive data, to ensure the privacy of the training data it is critical to identify and mitigate such unintended memorization. Federated Learning (FL) has emerged as a novel framework for large-scale distributed learning tasks. However, it differs in many aspects from the well-studied central learning setting where all the data is stored at the central server. In this paper, we initiate a formal study to understand the effect of different components of canonical FL on unintended memorization in trained models, comparing with the central learning setting. Our results show that several differing components of FL play an important role in reducing unintended memorization. Specifically, we observe that the clustering of data according to users---which happens by design in FL---has a significant effect in reducing such memorization, and using the method of Federated Averaging for training causes a further reduction. We also show that training with a strong user-level differential privacy guarantee results in models that exhibit the least amount of unintended memorization.

Shuang Song, Thomas Steinke, Om Thakkar et al.

We revisit the well-studied problem of differentially private empirical risk minimization (ERM). We show that for unconstrained convex generalized linear models (GLMs), one can obtain an excess empirical risk of $\tilde O\left(\sqrt{\texttt{rank}}/εn\right)$, where ${\texttt{rank}}$ is the rank of the feature matrix in the GLM problem, $n$ is the number of data samples, and $ε$ is the privacy parameter. This bound is attained via differentially private gradient descent (DP-GD). Furthermore, via the first lower bound for unconstrained private ERM, we show that our upper bound is tight. In sharp contrast to the constrained ERM setting, there is no dependence on the dimensionality of the ambient model space ($p$). (Notice that ${\texttt{rank}}\leq \min\{n, p\}$.) Besides, we obtain an analogous excess population risk bound which depends on ${\texttt{rank}}$ instead of $p$. For the smooth non-convex GLM setting (i.e., where the objective function is non-convex but preserves the GLM structure), we further show that DP-GD attains a dimension-independent convergence of $\tilde O\left(\sqrt{\texttt{rank}}/εn\right)$ to a first-order-stationary-point of the underlying objective. Finally, we show that for convex GLMs, a variant of DP-GD commonly used in practice (which involves clipping the individual gradients) also exhibits the same dimension-independent convergence to the minimum of a well-defined objective. To that end, we provide a structural lemma that characterizes the effect of clipping on the optimization profile of DP-GD.

Ryan Rogers, Aaron Roth, Adam Smith et al.

We design a general framework for answering adaptive statistical queries that focuses on providing explicit confidence intervals along with point estimates. Prior work in this area has either focused on providing tight confidence intervals for specific analyses, or providing general worst-case bounds for point estimates. Unfortunately, as we observe, these worst-case bounds are loose in many settings --- often not even beating simple baselines like sample splitting. Our main contribution is to design a framework for providing valid, instance-specific confidence intervals for point estimates that can be generated by heuristics. When paired with good heuristics, this method gives guarantees that are orders of magnitude better than the best worst-case bounds. We provide a Python library implementing our method.

Galen Andrew, Om Thakkar, H. Brendan McMahan et al.

Existing approaches for training neural networks with user-level differential privacy (e.g., DP Federated Averaging) in federated learning (FL) settings involve bounding the contribution of each user's model update by clipping it to some constant value. However there is no good a priori setting of the clipping norm across tasks and learning settings: the update norm distribution depends on the model architecture and loss, the amount of data on each device, the client learning rate, and possibly various other parameters. We propose a method wherein instead of a fixed clipping norm, one clips to a value at a specified quantile of the update norm distribution, where the value at the quantile is itself estimated online, with differential privacy. The method tracks the quantile closely, uses a negligible amount of privacy budget, is compatible with other federated learning technologies such as compression and secure aggregation, and has a straightforward joint DP analysis with DP-FedAvg. Experiments demonstrate that adaptive clipping to the median update norm works well across a range of realistic federated learning tasks, sometimes outperforming even the best fixed clip chosen in hindsight, and without the need to tune any clipping hyperparameter.

Raef Bassily, Om Thakkar, Abhradeep Thakurta

We design differentially private learning algorithms that are agnostic to the learning model. Our algorithms are interactive in nature, i.e., instead of outputting a model based on the training data, they provide predictions for a set of $m$ feature vectors that arrive online. We show that, for the feature vectors on which an ensemble of models (trained on random disjoint subsets of a dataset) makes consistent predictions, there is almost no-cost of privacy in generating accurate predictions for those feature vectors. To that end, we provide a novel coupling of the distance to instability framework with the sparse vector technique. We provide algorithms with formal privacy and utility guarantees for both binary/multi-class classification, and soft-label classification. For binary classification in the standard (agnostic) PAC model, we show how to bootstrap from our privately generated predictions to construct a computationally efficient private learner that outputs a final accurate hypothesis. Our construction - to the best of our knowledge - is the first computationally efficient construction for a label-private learner. We prove sample complexity upper bounds for this setting. As in non-private sample complexity bounds, the only relevant property of the given concept class is its VC dimension. For soft-label classification, our techniques are based on exploiting the stability properties of traditional learning algorithms, like stochastic gradient descent (SGD). We provide a new technique to boost the average-case stability properties of learning algorithms to strong (worst-case) stability properties, and then exploit them to obtain private classification algorithms. In the process, we also show that a large class of SGD methods satisfy average-case stability properties, in contrast to a smaller class of SGD methods that are uniformly stable as shown in prior work.

Prateek Jain, Om Thakkar, Abhradeep Thakurta

We provide the first provably joint differentially private algorithm with formal utility guarantees for the problem of user-level privacy-preserving collaborative filtering. Our algorithm is based on the Frank-Wolfe method, and it consistently estimates the underlying preference matrix as long as the number of users $m$ is $ω(n^{5/4})$, where $n$ is the number of items, and each user provides her preference for at least $\sqrt{n}$ randomly selected items. Along the way, we provide an optimal differentially private algorithm for singular vector computation, based on the celebrated Oja's method, that provides significant savings in terms of space and time while operating on sparse matrices. We also empirically evaluate our algorithm on a suite of datasets, and show that it consistently outperforms the state-of-the-art private algorithms.

Ryan Rogers, Aaron Roth, Adam Smith et al.

In this paper, we initiate a principled study of how the generalization properties of approximate differential privacy can be used to perform adaptive hypothesis testing, while giving statistically valid $p$-value corrections. We do this by observing that the guarantees of algorithms with bounded approximate max-information are sufficient to correct the $p$-values of adaptively chosen hypotheses, and then by proving that algorithms that satisfy $(ε,δ)$-differential privacy have bounded approximate max information when their inputs are drawn from a product distribution. This substantially extends the known connection between differential privacy and max-information, which previously was only known to hold for (pure) $(ε,0)$-differential privacy. It also extends our understanding of max-information as a partially unifying measure controlling the generalization properties of adaptive data analyses. We also show a lower bound, proving that (despite the strong composition properties of max-information), when data is drawn from a product distribution, $(ε,δ)$-differentially private algorithms can come first in a composition with other algorithms satisfying max-information bounds, but not necessarily second if the composition is required to itself satisfy a nontrivial max-information bound. This, in particular, implies that the connection between $(ε,δ)$-differential privacy and max-information holds only for inputs drawn from product distributions, unlike the connection between $(ε,0)$-differential privacy and max-information.