Sequential Data Poisoning in LLM Post-Training

LLM post-training proceeds through multiple stages, e.g., supervised fine-tuning (SFT) followed by reinforcement learning from human feedback (RLHF) or direct preference optimization (DPO), where each stage draws data from different, potentially untrusted sources. Existing literature assumes data poisoning attacks may occur at each training stage, but neglects the possibility of multiple attackers. To study the trustworthiness of the entire post-training pipeline, we propose the threat model of sequential data poisoning, where multiple adversaries separately poison the SFT and preference datasets. Under this threat model, we identify the single-attacker illusion: each adversary, evaluated in isolation, appears to pose a negligible threat. Yet when adversaries collaborate across stages, the true vulnerability is revealed. In the SFT $\to$ DPO pipeline, their contributions are additive: splitting a fixed poison budget across stages outperforms concentrating it in either stage alone. In the SFT $\to$ PPO pipeline, their contributions are complementary: neither SFT nor reward model poisoning succeeds individually, yet their combination does. These findings show that security analyses of individual post-training stages systematically underestimate compound vulnerabilities that emerge only from their interaction. Code is available at https://github.com/jcksanderson/sequential-poisoning.

Near-Optimal Pure Machine Unlearning for Smooth Strongly Convex Losses

Machine unlearning is motivated by legal and user-facing requirements to remove the influence of individuals' data from trained models, such as the right to be forgotten. Prior work has developed algorithms and error bounds for unlearning in smooth strongly convex stochastic optimization, but the fundamental statistical cost of unlearning has remained unclear. We nearly resolve this problem by proving upper and lower bounds on the excess population risk of approximate $\varepsilon$-unlearning; our bounds are tight up to a condition-number factor. For mean estimation over the unit ball, our upper and lower bounds match. The optimal rate is the usual statistical error plus an unlearning penalty that interpolates between the retraining-from-scratch rate and an exponentially smaller term as $\varepsilon/d$ grows, where $d$ is the dimension of the model. In particular, when $\varepsilon \gg d$, our $\varepsilon$-unlearning algorithm offers an exponential accuracy improvement over retraining the model from scratch and differentially private baselines. On the other hand, when $\varepsilon \le d$, retraining from scratch is optimal.

ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities?

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 $\varepsilon=100$.

Robust Statistical Estimators with Bounded Empirical Sensitivity

We introduce a new measure of robustness for statistical estimators, which we call \emph{empirical sensitivity}. An estimator $\hat θ$ has bounded empirical sensitivity if, with high probability over a dataset $X = (X_1, \dots, X_n) \sim \mathcal{D}^{\otimes n}$, for any dataset $Y$ obtained by modifying at most $ηn$ points in $X$, we have that $\hat θ(Y)$ is close to $\hat θ(X)$. We study bounds on this quantity for the prototypical problem of Gaussian mean estimation. We prove new lower bounds, showing that for any estimator $\hat μ$ which achieves an optimal $\ell_2$-error bound of $O\left(\sqrt{d/n}\right)$, the empirical sensitivity is at least $Ω\left(η+ \sqrt{ηd/n}\right)$. The two terms arise due to obstructions on the mean and variance (via an Efron-Stein argument) of such an estimator. We show that this bound is tight up to logarithmic factors, by employing recent results for robust empirical mean estimation.

DriftXpress: Faster Drifting Models via Projected RKHS Fields

Drifting Models have emerged as a new paradigm for one-step generative modeling, achieving strong image quality without iterative inference. The premise is to replace the iterative denoising process in diffusion models with a single evaluation of a generator. However, this creates a different trade-off: drifting reduces inference cost by moving much of the computation into training. We introduce DriftXpress, an accelerated formulation of drifting models based on projected RKHS fields. DriftXpress approximates the drifting kernel in a low-rank feature space. This preserves the attraction-repulsion structure of the original drifting field while reducing the cost of field evaluation. Across image-generation benchmarks, DriftXpress achieves comparable FID to standard drifting while reducing wall-clock training cost. These results show that the training-inference trade-off of drifting models can be pushed further without giving up their one-step inference advantage.

Rethinking Benchmarks for Differentially Private Image Classification

We revisit benchmarks for differentially private image classification. We suggest a comprehensive set of benchmarks, allowing researchers to evaluate techniques for differentially private machine learning in a variety of settings, including with and without additional data, in convex settings, and on a variety of qualitatively different datasets. We further test established techniques on these benchmarks in order to see which ideas remain effective in different settings. Finally, we create a publicly available leader board for the community to track progress in differentially private machine learning.

Query-Efficient Locally Private Hypothesis Selection via the Scheffe Graph

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.

Not All Samples Are Equal: Quantifying Instance-level Difficulty in Targeted Data Poisoning

Targeted data poisoning attacks pose an increasingly serious threat due to their ease of deployment and high success rates. These attacks aim to manipulate the prediction for a single test sample in classification models. Unlike indiscriminate attacks that aim to decrease overall test performance, targeted attacks present a unique threat to individual test instances. This threat model raises a fundamental question: what factors make certain test samples more susceptible to successful poisoning than others? We investigate how attack difficulty varies across different test instances and identify key characteristics that influence vulnerability. This paper introduces three predictive criteria for targeted data poisoning difficulty: ergodic prediction accuracy (analyzed through clean training dynamics), poison distance, and poison budget. Our experimental results demonstrate that these metrics effectively predict the varying difficulty of real-world targeted poisoning attacks across diverse scenarios, offering practitioners valuable insights for vulnerability assessment and understanding data poisoning attacks.

On the Learnability of Distribution Classes with Adaptive Adversaries

We consider the question of learnability of distribution classes in the presence of adaptive adversaries -- that is, adversaries capable of intercepting the samples requested by a learner and applying manipulations with full knowledge of the samples before passing it on to the learner. This stands in contrast to oblivious adversaries, who can only modify the underlying distribution the samples come from but not their i.i.d.\ nature. We formulate a general notion of learnability with respect to adaptive adversaries, taking into account the budget of the adversary. We show that learnability with respect to additive adaptive adversaries is a strictly stronger condition than learnability with respect to additive oblivious adversaries.

Optimal Differentially Private Sampling of Unbounded Gaussians

We provide the first $\widetilde{\mathcal{O}}\left(d\right)$-sample algorithm for sampling from unbounded Gaussian distributions under the constraint of $\left(\varepsilon, \delta\right)$-differential privacy. This is a quadratic improvement over previous results for the same problem, settling an open question of Ghazi, Hu, Kumar, and Manurangsi.