NeurIPS 2025 Papers

Machine unlearning algorithms aim to efficiently remove data from a model without retraining it from scratch, in order to remove corrupted or outdated data or respect a user's "right to be forgotten." Certified machine unlearning is a strong theoretical guarantee based on differential privacy that quantifies the extent to which an algorithm erases data from the model weights. In contrast to existing works in certified unlearning for convex or strongly convex loss functions, or nonconvex objectives with limiting assumptions, we propose the first, first-order, black-box (i.e., can be applied to models pretrained with vanilla gradient descent) algorithm for unlearning on general nonconvex loss functions, which unlearns by ``rewinding" to an earlier step during the learning process before performing gradient descent on the loss function of the retained data points. We prove $(\epsilon, \delta)$ certified unlearning and performance guarantees that establish the privacy-utility-complexity tradeoff of our algorithm, and we prove generalization guarantees for nonconvex functions that satisfy the Polyak-Lojasiewicz inequality. Finally, we demonstrate the superior performance of our algorithm compared to existing methods, within a new experimental framework that more accurately reflects unlearning user data in practice.

We study the problem of differentially private second moment estimation and present a new algorithm that achieve strong privacy-utility trade-offs even for worst-case inputs under subsamplability assumptions on the data. We call an input $(m,\alpha,\beta)$-subsamplable if a random subsample of size $m$ (or larger) preserves w.p $\geq 1-\beta$ the spectral structure of the original second moment matrix up to a multiplicative factor of $1\pm \alpha$. Building upon subsamplability, we give a recursive algorithmic framework similar to Kamath et al (2019) that abides zero-Concentrated Differential Privacy (zCDP) while preserving w.h.p the accuracy of the second moment estimation upto an arbitrary factor of $(1\pm\gamma)$. We then show how to apply our algorithm to approximate the second moment matrix of a distribution $\mathcal{D}$, even when a noticeable fraction of the input are outliers.

Differentially private (DP) machine learning often relies on the availability of public data for tasks like privacy-utility trade-off estimation, hyperparameter tuning, and pretraining. While public data assumptions may be reasonable in text and image data, they are less likely to hold for tabular data due to tabular data heterogeneity across domains. We propose leveraging powerful priors to address this limitation; specifically, we synthesize realistic tabular data directly from schema-level specifications -- such as variable names, types, and permissible ranges -- without ever accessing sensitive records. To that end, this work introduces the notion of ``surrogate'' public data -- datasets generated independently of sensitive data, which consume no privacy loss budget and are constructed solely from publicly available schema or metadata. Surrogate public data are intended to encode plausible statistical assumptions (informed by publicly available information) into a dataset with many downstream uses in private mechanisms. We automate the process of generating surrogate public data with large language models (LLMs); in particular, we propose two methods: direct record generation as CSV files, and automated structural causal model (SCM) construction for sampling records. Through extensive experiments, we demonstrate that surrogate public tabular data can effectively replace traditional public data when pretraining differentially private tabular classifiers. To a lesser extent, surrogate public data are also useful for hyperparameter tuning of DP synthetic data generators, and for estimating the privacy-utility tradeoff.

Large Language Models are typically trained on datasets collected from the web, which may inadvertently contain harmful or sensitive personal information. To address growing privacy concerns, unlearning methods have been proposed to remove the influence of specific data from trained models. Of these, exact unlearning---which retrains the model from scratch without the target data---is widely regarded as the gold standard for mitigating privacy risks in deployment. In this paper, we revisit this assumption in a practical deployment setting where both the pre- and post-unlearning logits API are exposed, such as in open-weight scenarios. Targeting this setting, we introduce a novel data extraction attack that leverages signals from the pre-unlearning model to guide the post-unlearning model, uncovering patterns that reflect the removed data distribution. Combining model guidance with a token filtering strategy, our attack significantly improves extraction success rates---doubling performance in some cases---across common benchmarks such as MUSE, TOFU, and WMDP. Furthermore, we demonstrate our attack's effectiveness on a simulated medical diagnosis dataset to highlight real-world privacy risks associated with exact unlearning. In light of our findings, which suggest that unlearning may, in a contradictory way, \textit{increase} the risk of privacy leakage during real-world deployments, we advocate for evaluation of unlearning methods to consider broader threat models that account not only for post-unlearning models but also for adversarial access to prior checkpoints. Code is publicly available at: https://github.com/Nicholas0228/unlearned_data_extraction_llm.

As major progress in LLM-based long-form text generation enables paradigms such as retrieval-augmented generation (RAG) and inference-time scaling, safely incorporating private information into the generation remains a critical open question. We present InvisibleInk, a highly scalable long-form text generation framework satisfying rigorous differential privacy guarantees with respect to the sensitive reference texts. It interprets sampling from the LLM's next-token-distribution as the exponential mechanism over the LLM logits with two innovations. First, we reduce the privacy cost by isolating and clipping only the sensitive information in the model logits (relative to the public logits). Second, we improve text quality by sampling without any privacy cost from a small superset of the top-$k$ private tokens. Empirical evaluations demonstrate a consistent $8\times$ (or more) reduction in computation cost over state-of-the-art baselines to generate long-form private text of the same utility across privacy levels. InvisibleInk is able to generate, for the first time, high-quality private long-form text at less than $4\text{-}8\times$ times the computation cost of non-private generation, paving the way for its practical use.

People have diverse privacy requirements. This is best modeled using a personalized local differential privacy model where each user privatizes their data using a possibly different privacy parameter. While the model of personalized local differential privacy is a natural and important one, prior work has failed to give meaningful error bounds. In this paper, we study the foundational sum/mean estimation problem under this model. We present two novel protocols that achieve strong error guarantees. The first gives a guarantee based on the radius of the data, suiting inputs that are centered around zero. The second extends the guarantee to the diameter of the data, capturing the case when the points are situated arbitrarily. Experimental results on both synthetic and real data show that our protocols significantly outperform existing methods in terms of accuracy while providing a strong level of privacy.

Communication cost and privacy are two major considerations in federated learning (FL). For communication cost, gradient compression by sketching the clients’ transmitted model updates is often used for reducing per‐round communication. For privacy, the Gaussian mechanism (GM), which consists of clipping updates and adding Gaussian noise, is commonly used to guarantee client‐level differential privacy. Existing literature on private FL analyzes privacy of sketching and GM in an isolated manner, illustrating that sketching provides privacy determined by the sketching dimension and that GM has to supply any additional desired privacy. In this paper, we introduce the Sketched Gaussian Mechanism (SGM), which directly combines sketching and the Gaussian mechanism for privacy. Using Rényi-DP tools, we present a joint analysis of SGM's overall privacy guarantee, which is significantly more flexible and sharper compared to isolated analysis of sketching and GM privacy. In particular, we prove that the privacy level of SGM for a fixed noise magnitude is proportional to $1/\sqrt{b}$, where $b$ is the sketching dimension, indicating that (for moderate $b$) SGM can provide much stronger privacy guarantees than the original GM under the same noise budget. We demonstrate the application of SGM to FL with either gradient descent or adaptive server optimizers, and establish theoretical results on optimization convergence, which exhibits only a logarithmic dependence on the number of parameters $d$. Experimental results confirm that at the same privacy level, SGM based FL is at least competitive with non‐sketching private FL variants and outperforms them in some settings. Moreover, using adaptive optimization at the server improves empirical performance while maintaining the privacy guarantees.

The "black-box service model" enables ML service providers to serve clients while keeping their intellectual property and client data confidential. Confidentiality is critical for delivering ML services legally and responsibly, but makes it difficult for outside parties to verify important model properties such as fairness. Existing methods that assess model fairness confidentially lack either (i) *reliability* because they certify fairness with respect to a static set of data, and therefore fail to guarantee fairness in the presence of distribution shift or service provider malfeasance; and/or (ii) *scalability* due to the computational overhead of confidentiality-preserving cryptographic primitives. We address these problems by introducing *online fairness certificates*, which verify that a model is fair with respect to data received by the service provider *online* during deployment. We then present OATH, a deployably efficient and scalable zero-knowledge proof protocol for confidential online group fairness certification. OATH exploits statistical properties of group fairness via a "cut-and-choose" style protocol, enabling scalability improvements over baselines.

Private large language model (LLM) inference based on cryptographic primitives offers a promising path towards privacy-preserving deep learning. However, existing frameworks only support dense LLMs like LLaMA-1 and struggle to scale to mixture-of-experts (MoE) architectures. The key challenge comes from securely evaluating the dynamic routing mechanism in MoE layers, which may reveal sensitive input information if not fully protected. In this paper, we propose CryptoMoE, the first framework that enables private, efficient, and accurate inference for MoE-based models. CryptoMoE balances expert loads to protect expert routing information and proposes novel protocols for secure expert dispatch and combine. CryptoMoE also develops a confidence-aware token selection strategy and a batch matrix multiplication protocol to improve accuracy and efficiency further. Extensive experiments on DeepSeekMoE-16.4B, OLMoE-6.9B, and QWenMoE-14.3B show that CryptoMoE achieves $2.8\sim3.5\times$ end-to-end latency reduction and $3\sim6\times$ communication reduction over a dense baseline with minimum accuracy loss. We also adapt CipherPrune (ICLR'25) for MoE inference and demonstrate CryptoMoE can reduce the communication by up to $4.3 \times$.

Given an undirected, weighted $n$-vertex graph $G = (V, E, w)$, a Gomory-Hu tree $T$ is a weighted tree on $V$ that preserves the Min-$s$-$t$-Cut between any pair of vertices $s, t \in V$. Finding cuts in graphs is a key primitive in problems such as bipartite matching, spectral and correlation clustering, and community detection. We design a differentially private (DP) algorithm that computes an approximate Gomory-Hu tree. Our algorithm is $\varepsilon$-DP, runs in polynomial time, and can be used to compute $s$-$t$ cuts that are $\tilde{O}(n/\varepsilon)$-additive approximations of the Min-$s$-$t$-Cuts in $G$ for all distinct $s, t \in V$ with high probability. Our error bound is essentially optimal, since [Dalirrooyfard, Mitrovic and Nevmyvaka, Neurips 2023] showed that privately outputting a single Min-$s$-$t$-Cut requires $\Omega(n)$ additive error even with $(\varepsilon, \delta)$-DP and allowing for multiplicative error. Prior to our work, the best additive error bounds for approximate all-pairs Min-$s$-$t$-Cuts were $O(n^{3/2}/\varepsilon)$ for $\varepsilon$-DP [Gupta, Roth, Ullman, TCC 2009] and $\tilde{O}(\sqrt{mn}/ \varepsilon)$ for $(\varepsilon, \delta)$-DP [Liu, Upadhyay and Zou, SODA 2024], both achieved by DP algorithms that preserve all cuts in the graph. To achieve our result, we develop an $\varepsilon$-DP algorithm for the Minimum Isolating Cuts problem with near-linear error, and introduce a novel privacy composition technique combining elements of both parallel and basic composition to handle `bounded overlap' computational branches in recursive algorithms, which maybe of independent interest.

Federated Learning (FL) enables collaborative model training across institutions without sharing raw data. However, gradient sharing still risks privacy leakage, such as gradient inversion attacks. Homomorphic Encryption (HE) can secure aggregation but often incurs prohibitive computational and communication overhead. Existing HE-based FL methods sit at two extremes: encrypting all gradients for full privacy at high cost, or partially encrypting gradients to save resources while exposing vulnerabilities. We present **DictPFL**, a practical framework that achieves full gradient protection with minimal overhead. DictPFL encrypts every transmitted gradient while keeping non-transmitted parameters local, preserving privacy without heavy computation. It introduces two key modules: **Decompose-for-Partial-Encrypt (DePE)**, which decomposes model weights into a static dictionary and an updatable lookup table—only the latter is encrypted and aggregated, while the static dictionary remains local and requires neither sharing nor encryption; and **Prune-for-Minimum-Encrypt (PrME)**, which applies encryption-aware pruning to minimize encrypted parameters via consistent, history-guided masks. Experiments show that DictPFL reduces communication cost by 402-748$\times$ and accelerates training by 28-65$\times$ compared to fully encrypted FL, while outperforming state-of-the-art selective encryption methods by 51-155$\times$ in overhead and 4-19$\times$ in speed. Remarkably, DictPFL’s runtime is within 2$\times$ of plaintext FL, demonstrating, for the first time, that HE-based private federated learning is practical for real-world deployment. The code is publicly available at https://github.com/UCF-ML-Research/DictPFL.

Recent advances in diffusion models have enabled high-quality synthesis of specific subjects, such as identities or objects. This capability, while unlocking new possibilities in content creation, also introduces significant privacy risks, as personalization techniques can be misused by malicious users to generate unauthorized images. Although several studies have attempted to counter this by generating adversarially perturbed samples designed to disrupt personalization, they rely on unrealistic assumptions and become ineffective in the presence of even a few clean images or under simple image transformations. To address these challenges, we shift the protection target from the images to the diffusion model itself to hinder the personalization of specific subjects, through our novel framework called $\textbf{A}$nti-$\textbf{P}$ersonalized $\textbf{D}$iffusion $\textbf{M}$odels ($\textbf{APDM}$). We first provide a theoretical analysis demonstrating that a naive approach of existing loss functions to diffusion models is inherently incapable of ensuring convergence for robust anti-personalization. Motivated by this finding, we introduce Direct Protective Optimization (DPO), a novel loss function that effectively disrupts subject personalization in the target model without compromising generative quality. Moreover, we propose a new dual-path optimization strategy, coined Learning to Protect (L2P). By alternating between personalization and protection paths, L2P simulates future personalization trajectories and adaptively reinforces protection at each step. Experimental results demonstrate that our framework outperforms existing methods, achieving state-of-the-art performance in preventing unauthorized personalization. The code is available at https://github.com/KU-VGI/APDM.

Differential privacy (DP) limits the impact of individual training data samples by bounding their gradient norms through clipping. Conventional clipping operations assign unequal scaling factors to sample gradients with different norms, leading to a direction mismatch between the true batch gradient and the aggregation of the clipped gradients. Applying a smaller but identical scaling factor to all sample gradients alleviates this direction mismatch; however, it intensifies the magnitude mismatch by excessively reducing the aggregation norm. This work proposes a novel clipping method, termed adaptive sigmoid (AdaSig), which uses a sigmoid function with an adjustable saturation slope to clip the sample gradients. The slope is adaptively adjusted during the training process to balance the trade-off between direction mismatch and magnitude mismatch, as the statistics of sample gradients evolve over the training iterations. Despite AdaSig’s adaptive nature, our convergence analysis demonstrates that differentially private stochastic gradient descent (DP-SGD) with AdaSig clipping retains the best-known convergence rate under non-convex loss functions. Evaluating AdaSig on sentence and image classification tasks across different datasets shows that it consistently improves learning performance compared with established clipping methods.

We consider the problem of estimating the correlation of two random variables $X$ and $Y$, where the pairs $(X,Y)$ are not observed together, but are instead separated co-ordinate-wise at two servers: server 1 contains all the $X$ observations, and server 2 contains the corresponding $Y$ observations. In this vertically distributed setting, we assume that each server has its own privacy constraints, owing to which they can only share suitably privatized statistics of their own component observations. We consider differing privacy budgets $(\varepsilon_1,\delta_1)$ and $(\varepsilon_2,\delta_2)$ for the two servers and determine the minimax optimal rates for correlation estimation allowing for both non-interactive and interactive mechanisms. We also provide correlation estimators that achieve these rates and further develop inference procedures, namely, confidence intervals, for the estimated correlations. Our results are characterized by an interesting rate in terms of the sample size $n$, $\varepsilon_1$, $\varepsilon_2$, which is strictly slower than the usual central privacy estimation rates. More interestingly, we find that the interactive mechanism is always better than its non-interactive counterpart whenever the two privacy budgets are different. Results from extensive numerical experiments support our theoretical findings.

We propose an algorithm with improved query-complexity for the problem of hypothesis selection under local differential privacy constraints. Given a set of $k$ probability distributions $Q$, we describe an algorithm that satisfies local differential privacy, performs $\tilde{O}(k^{3/2})$ non-adaptive queries to individuals who each have samples from a probability distribution $p$, and outputs a probability distribution from the set $Q$ which is nearly the closest to $p$. Previous algorithms required either $\Omega(k^2)$ queries or many rounds of interactive queries. Technically, we introduce a new object we dub the Scheff\'e graph, which captures structure of the differences between distributions in $Q$, and may be of more broad interest for hypothesis selection tasks.

Machine unlearning offers a practical alternative to avoid full model re-training by approximately removing the influence of specific user data. While existing methods certify unlearning via statistical indistinguishability from re-trained models, these guarantees do not naturally extend to model outputs when inputs are adversarially perturbed. In particular, slight perturbations of forget samples may still be correctly recognized by the unlearned model---even when a re-trained model fails to do so---revealing a novel privacy risk: information about the forget samples may persist in their local neighborhood. In this work, we formalize this vulnerability as residual knowledge and show that it is inevitable in high-dimensional settings. To mitigate this risk, we propose a fine-tuning strategy, named RURK, that penalizes the model’s ability to re-recognize perturbed forget samples. Experiments on vision benchmarks with deep neural networks demonstrate that residual knowledge is prevalent across existing unlearning methods and that our approach effectively prevents residual knowledge.

With increasing regulations on private data usage in AI systems, machine unlearning has emerged as a critical solution for selectively removing sensitive information from trained models while preserving their overall utility. While many existing unlearning methods rely on the *retain data* to mitigate the performance decline caused by forgetting, such data may not always be available (*retain-free*) in real-world scenarios. To address this challenge posed by retain-free unlearning, we introduce **RUAGO**, utilizing adversarial soft labels to mitigate over-unlearning and a generative model pretrained on out-of-distribution (OOD) data to effectively distill the original model’s knowledge. We introduce a progressive sampling strategy to incrementally increase synthetic data complexity, coupled with an inversion-based alignment step that ensures the synthetic data closely matches the original training distribution. Our extensive experiments on multiple benchmark datasets and architectures demonstrate that our approach consistently outperforms existing retain-free methods and achieves comparable or superior performance relative to retain-based approaches, demonstrating its effectiveness and practicality in real-world, data-constrained environments.

We study the ability of language models to reason about appropriate information disclosure - a central aspect of the evolving field of agentic privacy. Whereas previous works have focused on evaluating a model's ability to align with human decisions, we examine the role of ambiguity and missing context on model performance when making information-sharing decisions. We identify context ambiguity as a crucial barrier for high performance in privacy assessments. By designing Camber, a framework for context disambiguation, we show that model-generated decision rationales can reveal ambiguities and that systematically disambiguating context based on these rationales leads to significant accuracy improvements (up to 13.3% in precision and up to 22.3% in recall) as well as reductions in prompt sensitivity. Overall, our results indicate that approaches for context disambiguation are a promising way forward to enhance agentic privacy reasoning.

Recent methods for auditing the privacy of machine learning algorithms have improved computational efficiency by simultaneously intervening on multiple training examples in a single training run. Steinke et al. prove that one-run auditing indeed lower bounds the true privacy parameter of the audited algorithm, and give impressive empirical results. Their work leaves open the question of how precisely one-run auditing can uncover the true privacy parameter of an algorithm, and how that precision depends on the audited algorithm. In this work, we characterize the maximum achievable efficacy of one-run auditing and show that the key barrier to its efficacy is interference between the observable effects of different data elements. We present new conceptual approaches to minimize this barrier, towards improving the performance of one-run auditing of real machine learning algorithms.