Weiwei Kong

Peihan Liu, Lucas Rosenblatt, Weiwei Kong et al.

Differentially private (DP) text synthesis promises to unlock sensitive corpora for model training, but it remains unclear whether DP synthetic data transmits genuinely new knowledge and capabilities present only in those corpora. This is because existing evaluations rely on tasks that are nearly solvable without training, so strong benchmark performance does not establish that DP synthesis can substitute original data access. Thus, we introduce ContinuousBench, a continuously and automatically-regenerated benchmark that measures capability gain from DP synthetic text. Each quarter, a new release pairs a never-before-seen training corpus with a derived QA set, constructed to be: (1) unsolvable sans-corpus; and (2) learnable under DP, as the tested knowledge is supported by hundreds of independent records. Researchers produce DP synthetic data from the training corpus and run our standardized training and evaluation harness on their synthetic data to measure gains. We instantiate two tracks: Geminon, a procedurally-generated dataset about fictional creatures; and News, a stream of newly crawled public news articles. Although standard benchmarks are nearly saturated, on ContinuousBench we find that non-private synthesis transfers substantial knowledge from the original corpus, while state-of-the-art DP synthesis methods generally fail to do so, even at $\varepsilon=100$.

Kareem Amin, Alex Bie, Weiwei Kong et al.

We present an approach for generating differentially private synthetic text using large language models (LLMs), via private prediction. In the private prediction framework, we only require the output synthetic data to satisfy differential privacy guarantees. This is in contrast to approaches that train a generative model on potentially sensitive user-supplied source data and seek to ensure the model itself is safe to release. We prompt a pretrained LLM with source data, but ensure that next-token predictions are made with differential privacy guarantees. Previous work in this paradigm reported generating a small number of examples (<10) at reasonable privacy levels, an amount of data that is useful only for downstream in-context learning or prompting. In contrast, we make changes that allow us to generate thousands of high-quality synthetic data points, greatly expanding the set of potential applications. Our improvements come from an improved privacy analysis and a better private selection mechanism, which makes use of the equivalence between the softmax layer for sampling tokens in LLMs and the exponential mechanism. Furthermore, we introduce a novel use of public predictions via the sparse vector technique, in which we do not pay privacy costs for tokens that are predictable without sensitive data; we find this to be particularly effective for structured data.

Jasmine Bayrooti, Weiwei Kong, Natalia Ponomareva et al.

Generative models trained on sensitive image datasets risk memorizing and reproducing individual training examples, making strong privacy guarantees essential. While differential privacy (DP) provides a principled framework for such guarantees, standard DP finetuning (e.g., with DP-SGD) often results in severe degradation of image quality, particularly in high-frequency textures, due to the indiscriminate addition of noise across all model parameters. In this work, we propose a spectral DP framework based on the hypothesis that the most privacy-sensitive portions of an image are often low-frequency components in the wavelet space (e.g., facial features and object shapes) while high-frequency components are largely generic and public. Based on this hypothesis, we propose the following two-stage framework for DP image generation with coarse image intermediaries: (1) DP finetune an autoregressive spectral image tokenizer model on the low-resolution wavelet coefficients of the sensitive images, and (2) perform high-resolution upsampling using a publicly pretrained super-resolution model. By restricting the privacy budget to the global structures of the image in the first stage, and leveraging the post-processing property of DP for detail refinement, we achieve promising trade-offs between privacy and utility. Experiments on the MS-COCO and MM-CelebA-HQ datasets show that our method generates images with improved quality and style capture relative to other leading DP image frameworks.

Weiwei Kong, Mónica Ribero

Differentially-private stochastic gradient descent (DP-SGD) is a family of iterative machine learning training algorithms that privatize gradients to generate a sequence of differentially-private (DP) model parameters. It is also the standard tool used to train DP models in practice, even though most users are only interested in protecting the privacy of the final model. Tight DP accounting for the last iterate would minimize the amount of noise required while maintaining the same privacy guarantee and potentially increasing model utility. However, last-iterate accounting is challenging, and existing works require strong assumptions not satisfied by most implementations. These include assuming (i) the global sensitivity constant is known - to avoid gradient clipping; (ii) the loss function is Lipschitz or convex; and (iii) input batches are sampled randomly. In this work, we forego any unrealistic assumptions and provide privacy bounds for the most commonly used variant of DP-SGD, in which data is traversed cyclically, gradients are clipped, and only the last model is released. More specifically, we establish new Renyi differential privacy (RDP) upper bounds for the last iterate under realistic assumptions of small stepsize and Lipschitz smoothness of the loss function. Our general bounds also recover the special-case convex bounds when the weak-convexity parameter of the objective function approaches zero and no clipping is performed. The approach itself leverages optimal transport techniques for last iterate bounds, which is a nontrivial task when the data is traversed cyclically and the loss function is nonconvex.

Weiwei Kong, Andrés Muñoz Medina, Mónica Ribero

Unsupervised pre-training is a common step in developing computer vision models and large language models. In this setting, the absence of labels requires the use of similarity-based loss functions, such as contrastive loss, that favor minimizing the distance between similar inputs and maximizing the distance between distinct inputs. As privacy concerns mount, training these models using differential privacy has become more important. However, due to how inputs are generated for these losses, one of their undesirable properties is that their $L_2$ sensitivity grows with the batch size. This property is particularly disadvantageous for differentially private training methods, such as DP-SGD. To overcome this issue, we develop a new DP-SGD variant for similarity based loss functions -- in particular, the commonly-used contrastive loss -- that manipulates gradients of the objective function in a novel way to obtain a sensitivity of the summed gradient that is $O(1)$ for batch size $n$. We test our DP-SGD variant on some CIFAR-10 pre-training and CIFAR-100 finetuning tasks and show that, in both tasks, our method's performance comes close to that of a non-private model and generally outperforms DP-SGD applied directly to the contrastive loss.

Kareem Amin, Salman Avestimehr, Sara Babakniya et al.

Differentially private (DP) language model inference is an approach for generating private synthetic text. A sensitive input example is used to prompt an off-the-shelf large language model (LLM) to produce a similar example. Multiple examples can be aggregated together to formally satisfy the DP guarantee. Prior work creates inference batches by sampling sensitive inputs uniformly at random. We show that uniform sampling degrades the quality of privately generated text, especially when the sensitive examples concern heterogeneous topics. We remedy this problem by clustering the input data before selecting inference batches. Next, we observe that clustering also leads to more similar next-token predictions across inferences. We use this insight to introduce a new algorithm that aggregates next token statistics by privately computing medians instead of averages. This approach leverages the fact that the median has decreased local sensitivity when next token predictions are similar, allowing us to state a data-dependent and ex-post DP guarantee about the privacy properties of this algorithm. Finally, we demonstrate improvements in terms of representativeness metrics (e.g., MAUVE) as well as downstream task performance. We show that our method produces high-quality synthetic data at significantly lower privacy cost than a previous state-of-the-art method.

Kareem Amin, Sara Babakniya, Alex Bie et al.

Synthetically-generated data plays an increasingly larger role in training large language models. However, while synthetic data has been found to be useful, studies have also shown that without proper curation it can cause LLM performance to plateau, or even "collapse", after many training iterations. In this paper, we formalize this question and develop a theoretical framework to investigate how much curation is needed in order to ensure that LLM performance continually improves. We find that the requirements are nearly minimal. We describe a training procedure that converges to an optimal LLM even if almost all of the non-synthetic training data is of poor quality. Our analysis is inspired by boosting, a classic machine learning technique that leverages a very weak learning algorithm to produce an arbitrarily good classifier. Our training procedure subsumes many recently proposed methods for training LLMs on synthetic data, and thus our analysis sheds light on why they are successful, and also suggests opportunities for future improvement. We present experiments that validate our theory, and show that dynamically focusing labeling resources on the most challenging examples -- in much the same way that boosting focuses the efforts of the weak learner -- leads to improved performance.

Kareem Amin, Alex Bie, Weiwei Kong et al.

The prevalence and low cost of LLMs have led to a rise of synthetic content. From review sites to court documents, ``natural'' content has been contaminated by data points that appear similar to natural data, but are in fact LLM-generated. In this work we revisit fundamental learning theory questions in this, now ubiquitous, setting. We model this scenario as a sequence of learning tasks where the input is a mix of natural and synthetic data, and the learning algorithms are oblivious to the origin of any individual example. We study the possibilities and limitations of ERM in this setting. For the problem of estimating the mean of an arbitrary $d$-dimensional distribution, we find that while ERM converges to the true mean, it is outperformed by an algorithm that assigns non-uniform weights to examples from different generations of data. For the PAC learning setting, the disparity is even more stark. We find that ERM does not always converge to the true concept, echoing the model collapse literature. However, we show there are algorithms capable of learning the correct hypothesis for arbitrary VC classes and arbitrary amounts of contamination.