Machine Unlearning Fails to Remove Data Poisoning Attacks

We revisit the efficacy of several practical methods for approximate machine unlearning developed for large-scale deep learning. In addition to complying with data deletion requests, one often-cited potential application for unlearning methods is to remove the effects of training on poisoned data. We experimentally demonstrate that, while existing unlearning methods have been demonstrated to be effective in a number of evaluation settings (e.g., alleviating membership inference attacks), they fail to remove the effects of data poisoning, across a variety of types of poisoning attacks (indiscriminate, targeted, and a newly-introduced Gaussian poisoning attack) and models (image classifiers and LLMs); even when granted a relatively large compute budget. In order to precisely characterize unlearning efficacy, we introduce new evaluation metrics for unlearning based on data poisoning. Our results suggest that a broader perspective, including a wider variety of evaluations, is required to avoid a false sense of confidence in machine unlearning procedures for deep learning without provable guarantees. Moreover, while unlearning methods show some signs of being useful to efficiently remove poisoned datapoints without having to retrain, our work suggests that these methods are not yet "ready for prime time", and currently provide limited benefit over retraining.

Distribution Learnability and Robustness

We examine the relationship between learnability and robust (or agnostic) learnability for the problem of distribution learning. We show that, contrary to other learning settings (e.g., PAC learning of function classes), realizable learnability of a class of probability distributions does not imply its agnostic learnability. We go on to examine what type of data corruption can disrupt the learnability of a distribution class and what is such learnability robust against. We show that realizable learnability of a class of distributions implies its robust learnability with respect to only additive corruption, but not against subtractive corruption. We also explore related implications in the context of compression schemes and differentially private learnability.

Private Mean Estimation with Person-Level Differential Privacy

We study differentially private (DP) mean estimation in the case where each person holds multiple samples. Commonly referred to as the "user-level" setting, DP here requires the usual notion of distributional stability when all of a person's datapoints can be modified. Informally, if $n$ people each have $m$ samples from an unknown $d$-dimensional distribution with bounded $k$-th moments, we show that \[n = \tilde \Theta\left(\frac{d}{\alpha^2 m} + \frac{d }{ \alpha m^{1/2} \varepsilon} + \frac{d}{\alpha^{k/(k-1)} m \varepsilon} + \frac{d}{\varepsilon}\right)\] people are necessary and sufficient to estimate the mean up to distance $\alpha$ in $\ell_2$-norm under $\varepsilon$-differential privacy (and its common relaxations). In the multivariate setting, we give computationally efficient algorithms under approximate DP (with slightly degraded sample complexity) and computationally inefficient algorithms under pure DP, and our nearly matching lower bounds hold for the most permissive case of approximate DP. Our computationally efficient estimators are based on the well known noisy-clipped-mean approach, but the analysis for our setting requires new bounds on the tails of sums of independent, vector-valued, bounded-moments random variables, and a new argument for bounding the bias introduced by clipping.

Avoiding Pitfalls for Privacy Accounting of Subsampled Mechanisms under Composition

We consider the problem of computing tight privacy guarantees for the composition of subsampled differentially private mechanisms. Recent algorithms can numerically compute the privacy parameters to arbitrary precision but must be carefully applied. Our main contribution is to address two common points of confusion. First, some privacy accountants assume that the privacy guarantees for the composition of a subsampled mechanism are determined by self-composing the worst-case datasets for the uncomposed mechanism. We show that this is not true in general. Second, Poisson subsampling is sometimes assumed to have similar privacy guarantees compared to sampling without replacement. We show that the privacy guarantees may in fact differ significantly between the two sampling schemes. In particular, we give an example of hyperparameters that result in $\varepsilon \approx 1$ for Poisson subsampling and $\varepsilon > 10$ for sampling without replacement. This occurs for some parameters that could realistically be chosen for DP-SGD.

Differentially Private Post-Processing for Fair Regression

This paper describes a differentially private post-processing algorithm for learning fair regressors satisfying statistical parity, addressing privacy concerns of machine learning models trained on sensitive data, as well as fairness concerns of their potential to propagate historical biases. Our algorithm can be applied to post-process any given regressor to improve fairness by remapping its outputs. It consists of three steps: first, the output distributions are estimated privately via histogram density estimation and the Laplace mechanism, then their Wasserstein barycenter is computed, and the optimal transports to the barycenter are used for post-processing to satisfy fairness. We analyze the sample complexity of our algorithm and provide fairness guarantee, revealing a trade-off between the statistical bias and variance induced from the choice of the number of bins in the histogram, in which using less bins always favors fairness at the expense of error.

Disguised Copyright Infringement of Latent Diffusion Models

Copyright infringement may occur when a generative model produces samples substantially similar to some copyrighted data that it had access to during the training phase. The notion of access usually refers to including copyrighted samples directly in the training dataset, which one may inspect to identify an infringement. We argue that such visual auditing largely overlooks a concealed copyright infringement, where one constructs a disguise that looks drastically different from the copyrighted sample yet still induces the effect of training Latent Diffusion Models on it. Such disguises only require indirect access to the copyrighted material and cannot be visually distinguished, thus easily circumventing the current auditing tools. In this paper, we provide a better understanding of such disguised copyright infringement by uncovering the disguises generation algorithm, the revelation of the disguises, and importantly, how to detect them to augment the existing toolbox. Additionally, we introduce a broader notion of acknowledgment for comprehending such indirect access.

Indiscriminate Data Poisoning Attacks on Pre-trained Feature Extractors

Machine learning models have achieved great success in supervised learning tasks for end-to-end training, which requires a large amount of labeled data that is not always feasible. Recently, many practitioners have shifted to self-supervised learning methods that utilize cheap unlabeled data to learn a general feature extractor via pre-training, which can be further applied to personalized downstream tasks by simply training an additional linear layer with limited labeled data. However, such a process may also raise concerns regarding data poisoning attacks. For instance, indiscriminate data poisoning attacks, which aim to decrease model utility by injecting a small number of poisoned data into the training set, pose a security risk to machine learning models, but have only been studied for end-to-end supervised learning. In this paper, we extend the exploration of the threat of indiscriminate attacks on downstream tasks that apply pre-trained feature extractors. Specifically, we propose two types of attacks: (1) the input space attacks, where we modify existing attacks to directly craft poisoned data in the input space. However, due to the difficulty of optimization under constraints, we further propose (2) the feature targeted attacks, where we mitigate the challenge with three stages, firstly acquiring target parameters for the linear head; secondly finding poisoned features by treating the learned feature representations as a dataset; and thirdly inverting the poisoned features back to the input space. Our experiments examine such attacks in popular downstream tasks of fine-tuning on the same dataset and transfer learning that considers domain adaptation. Empirical results reveal that transfer learning is more vulnerable to our attacks. Additionally, input space attacks are a strong threat if no countermeasures are posed, but are otherwise weaker than feature targeted attacks.

Not All Learnable Distribution Classes are Privately Learnable

We give an example of a class of distributions that is learnable in total variation distance with a finite number of samples, but not learnable under $(\varepsilon, \delta)$-differential privacy. This refutes a conjecture of Ashtiani.

Private Distribution Learning with Public Data: The View from Sample Compression

We study the problem of private distribution learning with access to public data. In this setup, which we refer to as public-private learning, the learner is given public and private samples drawn from an unknown distribution $p$ belonging to a class $\mathcal Q$, with the goal of outputting an estimate of $p$ while adhering to privacy constraints (here, pure differential privacy) only with respect to the private samples. We show that the public-private learnability of a class $\mathcal Q$ is connected to the existence of a sample compression scheme for $\mathcal Q$, as well as to an intermediate notion we refer to as list learning. Leveraging this connection: (1) approximately recovers previous results on Gaussians over $\mathbb R^d$; and (2) leads to new ones, including sample complexity upper bounds for arbitrary $k$-mixtures of Gaussians over $\mathbb R^d$, results for agnostic and distribution-shift resistant learners, as well as closure properties for public-private learnability under taking mixtures and products of distributions. Finally, via the connection to list learning, we show that for Gaussians in $\mathbb R^d$, at least $d$ public samples are necessary for private learnability, which is close to the known upper bound of $d+1$ public samples.

Exploring the Limits of Indiscriminate Data Poisoning Attacks

Indiscriminate data poisoning attacks aim to decrease a model's test accuracy by injecting a small amount of corrupted training data. Despite significant interest, existing attacks remain relatively ineffective against modern machine learning (ML) architectures. In this work, we introduce the notion of model poisonability as a technical tool to explore the intrinsic limits of data poisoning attacks. We derive an easily computable threshold to establish and quantify a surprising phase transition phenomenon among popular ML models: data poisoning attacks become effective only when the poisoning ratio exceeds our threshold. Building on existing parameter corruption attacks and refining the Gradient Canceling attack, we perform extensive experiments to confirm our theoretical findings, test the predictability of our transition threshold, and significantly improve existing data poisoning baselines over a range of datasets and models. Our work highlights the critical role played by the poisoning ratio, and sheds new insights on existing empirical results, attacks and mitigation strategies in data poisoning.