James Flemings

James Flemings, Ren Yi, Octavian Suciu et al.

Personal large language model (LLM) agents increasingly perform tasks that require access to user data, raising concerns about appropriate data disclosure. We show that relying solely on LLMs to make data-sharing decisions is insufficient. Prompting LLMs with general privacy norms fails to capture individual users' privacy preferences, while providing prior user data-sharing decisions through in-context learning (ICL) leads to unreliable and opaque reasoning. To address these limitations, we propose ARIEL (Agentic Reasoning with Individualized Entailment Logic), a framework that combines LLMs with rule-based logic to enable structured, personalized privacy reasoning. The core mechanism of ARIEL determines whether a user's prior decision on a data-sharing request $\textit{logically entails}$ the same decision for a new request. Experimental evaluations using advanced models and public datasets show that ARIEL reduces the F1 error rate for appropriate judgments by $\textbf{40.6%}$ compared to standard ICL-based reasoning, indicating that ARIEL is effective at correctly judging requests where the user would approve data sharing. These results demonstrate that integrating LLMs with logical entailment provides an effective and interpretable approach for automating personalized privacy decisions.

James Flemings, Murali Annavaram

Large Language models (LLMs) are achieving state-of-the-art performance in many different downstream tasks. However, the increasing urgency of data privacy puts pressure on practitioners to train LLMs with Differential Privacy (DP) on private data. Concurrently, the exponential growth in parameter size of LLMs necessitates model compression before deployment of LLMs on resource-constrained devices or latency-sensitive applications. Differential privacy and model compression generally must trade off utility loss to achieve their objectives. Moreover, simultaneously applying both schemes can compound the utility degradation. To this end, we propose DistilDP: a novel differentially private knowledge distillation algorithm that exploits synthetic data generated by a differentially private teacher LLM. The knowledge of a teacher LLM is transferred onto the student in two ways: one way from the synthetic data itself -- the hard labels, and the other way by the output distribution of the teacher evaluated on the synthetic data -- the soft labels. Furthermore, if the teacher and student share a similar architectural structure, we can further distill knowledge by aligning the hidden representations between both. Our experimental results demonstrate that DistilDP can substantially improve the utility over existing baselines, at least $9.0$ PPL on the Big Patent dataset, with strong privacy parameters, $ε=2$. These promising results progress privacy-preserving compression of autoregressive LLMs. Our code can be accessed here: https://github.com/james-flemings/dp_compress.

Tingting Tang, James Flemings, Yongqin Wang et al.

Retrieval-augmented generation (RAG) is a widely used framework for reducing hallucinations in large language models (LLMs) on domain-specific tasks by retrieving relevant documents from a database to support accurate responses. However, when the database contains sensitive corpora, such as medical records or legal documents, RAG poses serious privacy risks by potentially exposing private information through its outputs. Prior work has demonstrated that one can practically craft adversarial prompts that force an LLM to regurgitate the augmented contexts. A promising direction is to integrate differential privacy (DP), a privacy notion that offers strong formal guarantees, into RAG systems. However, naively applying DP mechanisms into existing systems often leads to significant utility degradation. Particularly for RAG systems, DP can reduce the usefulness of the augmented contexts leading to increase risk of hallucination from the LLMs. Motivated by these challenges, we present DP-KSA, a novel privacy-preserving RAG algorithm that integrates DP using the propose-test-release paradigm. DP-KSA follows from a key observation that most question-answering (QA) queries can be sufficiently answered with a few keywords. Hence, DP-KSA first obtains an ensemble of relevant contexts, each of which will be used to generate a response from an LLM. We utilize these responses to obtain the most frequent keywords in a differentially private manner. Lastly, the keywords are augmented into the prompt for the final output. This approach effectively compresses the semantic space while preserving both utility and privacy. We formally show that DP-KSA provides formal DP guarantees on the generated output with respect to the RAG database. We evaluate DP-KSA on two QA benchmarks using three instruction-tuned LLMs, and our empirical results demonstrate that DP-KSA achieves a strong privacy-utility tradeoff.

James Flemings, Murali Annavaram

Large language models (LLMs) are increasingly used to simulate human behavior, but their ability to simulate $individual$ privacy decisions is not well understood. In this paper, we address the problem of evaluating whether a core set of user persona attributes can drive LLMs to simulate individual-level privacy behavior. We introduce PrivacySIM, an evaluation suite that benchmarks LLM simulation of user privacy behavior against the ground-truth responses of 1,000 users. These users are drawn from five published user studies on privacy spanning LLM healthcare consultations, conversational agents, and chatbots. Drawing on these user studies, we hypothesize three persona facets as plausible predictors of privacy decision-making: demographics, previous experiences, and stated privacy attitudes. We condition nine frontier LLMs on subsets of these three facets and measure how often each model's response to a data-sharing scenario matches the user's actual response. Our findings show that (1) privacy persona conditioning consistently improves simulation quality over no-persona conditioning, but even the strongest model (40.4\% accuracy) remains far from faithfully simulating individual privacy decisions. (2) A user's stated privacy attitudes alone may not be the best predictor because they often diverge from the user's actual privacy behavior. (3) Users with high AI/chatbot experience but low stated privacy attitudes are the most challenging to simulate. PrivacySIM is a first step toward understanding and improving the capabilities of LLMs to simulate user privacy decisions. We release PrivacySIM to enable further evaluation of LLM privacy simulation.

Johnny Tian-Zheng Wei, Ameya Godbole, Mohammad Aflah Khan et al.

We present Hubble, a suite of fully open-source large language models (LLMs) for the scientific study of LLM memorization. Hubble models come in standard and perturbed variants: standard models are pretrained on a large English corpus, and perturbed models are trained in the same way but with controlled insertion of text (e.g., book passages, biographies, and test sets) designed to emulate key memorization risks. Our core release includes 8 models -- standard and perturbed models with 1B or 8B parameters, pretrained on 100B or 500B tokens -- establishing that memorization risks are determined by the frequency of sensitive data relative to size of the training corpus (i.e., a password appearing once in a smaller corpus is memorized better than the same password in a larger corpus). Our release also includes 6 perturbed models with text inserted at different pretraining phases, showing that sensitive data without continued exposure can be forgotten. These findings suggest two best practices for addressing memorization risks: to dilute sensitive data by increasing the size of the training corpus, and to order sensitive data to appear earlier in training. Beyond these general empirical findings, Hubble enables a broad range of memorization research; for example, analyzing the biographies reveals how readily different types of private information are memorized. We also demonstrate that the randomized insertions in Hubble make it an ideal testbed for membership inference and machine unlearning, and invite the community to further explore, benchmark, and build upon our work.

Mohammad Aflah Khan, Ameya Godbole, Johnny Tian-Zheng Wei et al.

Understanding the relationship between training data and model behavior during pretraining is crucial, but existing workflows make this process cumbersome, fragmented, and often inaccessible to researchers. We present TokenSmith, an open-source library for interactive editing, inspection, and analysis of datasets used in Megatron-style pretraining frameworks such as GPT-NeoX, Megatron, and NVIDIA NeMo. TokenSmith supports a wide range of operations including searching, viewing, ingesting, exporting, inspecting, and sampling data, all accessible through a simple user interface and a modular backend. It also enables structured editing of pretraining data without requiring changes to training code, simplifying dataset debugging, validation, and experimentation. TokenSmith is designed as a plug-and-play addition to existing large language model pretraining workflows, thereby democratizing access to production-grade dataset tooling. TokenSmith is hosted on GitHub, with accompanying documentation, tutorials, and a demonstration video (available on YouTube).

James Flemings, Meisam Razaviyayn, Murali Annavaram

Ensuring the privacy of Large Language Models (LLMs) is becoming increasingly important. The most widely adopted technique to accomplish this is DP-SGD, which trains a model to guarantee Differential Privacy (DP). However, DP-SGD overestimates an adversary's capabilities in having white box access to the model and, as a result, causes longer training times and larger memory usage than SGD. On the other hand, commercial LLM deployments are predominantly cloud-based; hence, adversarial access to LLMs is black-box. Motivated by these observations, we present Private Mixing of Ensemble Distributions (PMixED): a private prediction protocol for next-token prediction that utilizes the inherent stochasticity of next-token sampling and a public model to achieve Differential Privacy. We formalize this by introducing RD-mollifers which project each of the model's output distribution from an ensemble of fine-tuned LLMs onto a set around a public LLM's output distribution, then average the projected distributions and sample from it. Unlike DP-SGD which needs to consider the model architecture during training, PMixED is model agnostic, which makes PMixED a very appealing solution for current deployments. Our results show that PMixED achieves a stronger privacy guarantee than sample-level privacy and outperforms DP-SGD for privacy $ε= 8$ on large-scale datasets. Thus, PMixED offers a practical alternative to DP training methods for achieving strong generative utility without compromising privacy.

Alex Mulrooney, Devansh Gupta, James Flemings et al.

Differential privacy (DP) protects sensitive data during neural network training, but standard methods like DP-Adam suffer from high memory overhead due to per-sample gradient clipping, limiting scalability. We introduce DP-GRAPE (Gradient RAndom ProjEction), a DP training method that significantly reduces memory usage while maintaining utility on par with first-order DP approaches. Rather than directly applying DP to GaLore, DP-GRAPE introduces three key modifications: (1) gradients are privatized after projection, (2) random Gaussian matrices replace SVD-based subspaces, and (3) projection is applied during backpropagation. These contributions eliminate the need for costly SVD computations, enable substantial memory savings, and lead to improved utility. Despite operating in lower-dimensional subspaces, our theoretical analysis shows that DP-GRAPE achieves a privacy-utility trade-off comparable to DP-SGD. Our extensive empirical experiments show that DP-GRAPE can reduce the memory footprint of DP training without sacrificing accuracy or training time. In particular, DP-GRAPE reduces memory usage by over 63% when pre-training Vision Transformers and over 70% when fine-tuning RoBERTa-Large as compared to DP-Adam, while achieving similar performance. We further demonstrate that DP-GRAPE scales to fine-tuning large models such as OPT with up to 6.7 billion parameters.

James Flemings, Haosheng Gan, Hongyi Li et al.

In-context learning (ICL) has shown promising improvement in downstream task adaptation of LLMs by augmenting prompts with relevant input-output examples (demonstrations). However, the ICL demonstrations can contain privacy-sensitive information, which can be leaked and/or regurgitated by the LLM output. Differential Privacy (DP), a widely adopted privacy safeguard, has emerged to mitigate this privacy leakage, with recent work demonstrating strong privacy-utility tradeoffs in classification tasks for ICL. However, generation tasks for ICL are challenging due to the high-dimensional output space of open-ended generation. To this end, we propose $\texttt{dps-mozo}$, Differentially Private Sampling by Mixing One-shot with Zero-shot Outputs, a decoding framework that generates DP text by sampling from the product of multiple one-shot outputs mixed with a zero-shot output. This mixing effectively reduces the amount of information that can be leaked by each demonstration. By utilizing the inherent randomness in sampling from the mixed distributions, we can achieve DP without adding noise, thereby improving the privacy-utility tradeoff. Our experimental evaluations show $\texttt{dps-mozo}$ can achieve a strong privacy guarantee, $ε=2$, with minimal utility degradation compared to non-private few-shot learning, $\textbf{0.3}$% ROUGE-L F1 score decrease on the SAMSum dataset with Gemma 2 2B.