Open benchmarks are essential for evaluating and advancing large language models, offering reproducibility and transparency.However, their accessibility makes them likely targets of test set contamination.In this work, we introduce DyePack, a framework that leverages backdoor attacks to identify models that used benchmark test sets during training, without requiring access to the loss, logits, or any internal details of the model.Like how banks mix dye packs with their money to mark robbers, DyePack mixes backdoor samples with the test data to flag models that trained on it.We propose a principled design incorporating multiple backdoors with stochastic targets, enabling exact false positive rate (FPR) computation when flagging every model.This provably prevents false accusations while providing strong evidence for every detected case of contamination.We evaluate DyePack on five models across three datasets, covering both multiplechoice and open-ended generation tasks.For multiple-choice questions, it successfully detects all contaminated models with guaranteed FPRs as low as 0.000073% on MMLU-Pro and 0.000017% on Big-Bench-Hard using eight backdoors.For open-ended generation tasks, it generalizes well and identifies all contaminated models on Alpaca with a guaranteed false positive rate of just 0.127% using six backdoors.

Kazem Faghih, Wenxiao Wang, Yize Cheng, Siddhant Bharti, Gaurang Sriramanan, Sriram Balasubramanian, Parsa Hosseini, Soheil Feizi. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

In light of recent advancements in generative AI models, it has become essential to distinguish genuine content from AI-generated one to prevent the malicious usage of fake materials as authentic ones and vice versa. Various techniques have been introduced for identifying AI-generated images, with watermarking emerging as a promising approach. In this paper, we analyze the robustness of various AI-image detectors including watermarking and classifier-based deepfake detectors. For watermarking methods that introduce subtle image perturbations (i.e., low perturbation budget methods), we reveal a fundamental trade-off between the evasion error rate (i.e., the fraction of watermarked images detected as non-watermarked ones) and the spoofing error rate (i.e., the fraction of non-watermarked images detected as watermarked ones) upon an application of a diffusion purification attack. In this regime, we also empirically show that diffusion purification effectively removes watermarks with minimal changes to images. For high perturbation watermarking methods where notable changes are applied to images, the diffusion purification attack is not effective. In this case, we develop a model substitution adversarial attack that can successfully remove watermarks. Moreover, we show that watermarking methods are vulnerable to spoofing attacks where the attacker aims to have real images (potentially obscene) identified as watermarked ones, damaging the reputation of the developers. In particular, by just having black-box access to the watermarking method, we show that one can generate a watermarked noise image which can be added to the real images to have them falsely flagged as watermarked ones. Finally, we extend our theory to characterize a fundamental trade-off between the robustness and reliability of classifier-based deep fake detectors and demonstrate it through experiments. Code is available at https://github.com/mehrdadsaberi/watermark_robustness.

Machine Learning (ML) models have been utilized for malware detection for over two decades. Consequently, this ignited an ongoing arms race between malware authors and antivirus systems, compelling researchers to propose defenses for malware-detection models against evasion attacks. However, most if not all existing defenses against evasion attacks suffer from sizable performance degradation and/or can defend against only specific attacks, which makes them less practical in real-world settings. In this work, we develop a certified defense, DRSM (De-Randomized Smoothed MalConv), by redesigning the *de-randomized smoothing* technique for the domain of malware detection. Specifically, we propose a *window ablation* scheme to provably limit the impact of adversarial bytes while maximally preserving local structures of the executables. After showing how DRSM is theoretically robust against attacks with contiguous adversarial bytes, we verify its performance and certified robustness experimentally, where we observe only marginal accuracy drops as the cost of robustness. To our knowledge, we are the first to offer certified robustness in the realm of static detection of malware executables. More surprisingly, through evaluating DRSM against $9$ empirical attacks of different types, we observe that the proposed defense is empirically robust to some extent against a diverse set of attacks, some of which even fall out of the scope of its original threat model. In addition, we collected $15.5K$ recent benign raw executables from diverse sources, which will be made public as a dataset called PACE (Publicly Accessible Collection(s) of Executables) to alleviate the scarcity of publicly available benign datasets for studying malware detection and provide future research with more representative data of the time. Our code and dataset are available at - https://github.com/ShoumikSaha/DRSM

Data poisoning considers cases when an adversary manipulates the behavior of machine learning algorithms through malicious training data. Existing threat models of data poisoning center around a single metric, the number of poisoned samples. In consequence, if attackers can poison more samples than expected with affordable overhead, as in many practical scenarios, they may be able to render existing defenses ineffective in a short time. To address this issue, we leverage timestamps denoting the birth dates of data, which are often available but neglected in the past. Benefiting from these timestamps, we propose a temporal threat model of data poisoning with two novel metrics, earliness and duration, which respectively measure how long an attack started in advance and how long an attack lasted. Using these metrics, we define the notions of temporal robustness against data poisoning, providing a meaningful sense of protection even with unbounded amounts of poisoned samples when the attacks are temporally bounded. We present a benchmark with an evaluation protocol simulating continuous data collection and periodic deployments of updated models, thus enabling empirical evaluation of temporal robustness. Lastly, we develop and also empirically verify a baseline defense, namely temporal aggregation, offering provable temporal robustness and highlighting the potential of our temporal threat model for data poisoning.

We present a simple but effective method to measure and mitigate model biases caused by reliance on spurious cues. Instead of requiring costly changes to one's data or model training, our method better utilizes the data one already has by sorting them. Specifically, we rank images within their classes based on spuriosity (the degree to which common spurious cues are present), proxied via deep neural features of an interpretable network. With spuriosity rankings, it is easy to identify minority subpopulations (i.e. low spuriosity images) and assess model bias as the gap in accuracy between high and low spuriosity images. One can even efficiently remove a model's bias at little cost to accuracy by finetuning its classification head on low spuriosity images, resulting in fairer treatment of samples regardless of spuriosity. We demonstrate our method on ImageNet, annotating $5000$ class-feature dependencies ($630$ of which we find to be spurious) and generating a dataset of $325k$ soft segmentations for these features along the way. Having computed spuriosity rankings via the identified spurious neural features, we assess biases for $89$ diverse models and find that class-wise biases are highly correlated across models. Our results suggest that model bias due to spurious feature reliance is influenced far more by what the model is trained on than how it is trained.

Data poisoning attacks aim at manipulating model behaviors through distorting training data. Previously, an aggregation-based certified defense, Deep Partition Aggregation (DPA), was proposed to mitigate this threat. DPA predicts through an aggregation of base classifiers trained on disjoint subsets of data, thus restricting its sensitivity to dataset distortions. In this work, we propose an improved certified defense against general poisoning attacks, namely Finite Aggregation. In contrast to DPA, which directly splits the training set into disjoint subsets, our method first splits the training set into smaller disjoint subsets and then combines duplicates of them to build larger (but not disjoint) subsets for training base classifiers. This reduces the worst-case impacts of poison samples and thus improves certified robustness bounds. In addition, we offer an alternative view of our method, bridging the designs of deterministic and stochastic aggregation-based certified defenses. Empirically, our proposed Finite Aggregation consistently improves certificates on MNIST, CIFAR-10, and GTSRB, boosting certified fractions by up to 3.05%, 3.87% and 4.77%, respectively, while keeping the same clean accuracies as DPA’s, effectively establishing a new state of the art in (pointwise) certified robustness against data poisoning.

Data poisoning considers an adversary that distorts the training set of machine learning algorithms for malicious purposes. In this work, we bring to light one conjecture regarding the fundamentals of data poisoning, which we call the Lethal Dose Conjecture. The conjecture states: If $n$ clean training samples are needed for accurate predictions, then in a size-$N$ training set, only $\Theta(N/n)$ poisoned samples can be tolerated while ensuring accuracy. Theoretically, we verify this conjecture in multiple cases. We also offer a more general perspective of this conjecture through distribution discrimination. Deep Partition Aggregation (DPA) and its extension, Finite Aggregation (FA) are recent approaches for provable defenses against data poisoning, where they predict through the majority vote of many base models trained from different subsets of training set using a given learner. The conjecture implies that both DPA and FA are (asymptotically) optimal---if we have the most data-efficient learner, they can turn it into one of the most robust defenses against data poisoning. This outlines a practical approach to developing stronger defenses against poisoning via finding data-efficient learners. Empirically, as a proof of concept, we show that by simply using different data augmentations for base learners, we can respectively double and triple the certified robustness of DPA on CIFAR-10 and GTSRB without sacrificing accuracy.

Training deep neural networks from scratch could be computationally expensive and requires a lot of training data. Recent work has explored different watermarking techniques to protect the pre-trained deep neural networks from potential copyright infringements. However, these techniques could be vulnerable to watermark removal attacks. In this work, we propose REFIT, a unified watermark removal framework based on fine-tuning, which does not rely on the knowledge of the watermarks, and is effective against a wide range of watermarking schemes. In particular, we conduct a comprehensive study of a realistic attack scenario where the adversary has limited training data, which has not been emphasized in prior work on attacks against watermarking schemes. To effectively remove the watermarks without compromising the model functionality under this weak threat model, we propose two techniques that are incorporated into our fine-tuning framework: (1) an adaption of the elastic weight consolidation (EWC) algorithm, which is originally proposed for mitigating the catastrophic forgetting phenomenon; and (2) unlabeled data augmentation (AU), where we leverage auxiliary unlabeled data from other sources. Our extensive evaluation shows the effectiveness of REFIT against diverse watermark embedding schemes. In particular, both EWC and AU significantly decrease the amount of labeled training data needed for effective watermark removal, and the unlabeled data samples used for AU do not necessarily need to be drawn from the same distribution as the benign data for model evaluation. The experimental results demonstrate that our fine-tuning based watermark removal attacks could pose real threats to the copyright of pre-trained models, and thus highlight the importance of further investigating the watermarking problem and proposing more robust watermark embedding schemes against the attacks.

In self-supervised representation learning, a common idea behind most of the state-of-the-art approaches is to enforce the robustness of the representations to predefined augmentations. A potential issue of this idea is the existence of completely collapsed solutions (i.e., constant features), which are typically avoided implicitly by carefully chosen implementation details. In this work, we study a relatively concise framework containing the most common components from recent approaches. We verify the existence of complete collapse and discover another reachable collapse pattern that is usually overlooked, namely dimensional collapse. We connect dimensional collapse with strong correlations between axes and consider such connection as a strong motivation for feature decorrelation (i.e., standardizing the covariance matrix). The gains from feature decorrelation are verified empirically to highlight the importance and the potential of this insight.

Abstract Deep learning techniques have achieved remarkable performance in wide-ranging tasks. However, when trained on privacy-sensitive datasets, the model parameters may expose private information in training data. Prior attempts for differentially private training, although offering rigorous privacy guarantees, lead to much lower model performance than the non-private ones. Besides, different runs of the same training algorithm produce models with large performance variance. To address these issues, we propose DPlis– Differentially Private Learning wIth Smoothing. The core idea of DPlis is to construct a smooth loss function that favors noise-resilient models lying in large flat regions of the loss landscape. We provide theoretical justification for the utility improvements of DPlis. Extensive experiments also demonstrate that DPlis can effectively boost model quality and training stability under a given privacy budget.

This paper studies model-inversion attacks, in which the access to a model is abused to infer information about the training data. Since its first introduction by~\cite{fredrikson2014privacy}, such attacks have raised serious concerns given that training data usually contain privacy sensitive information. Thus far, successful model-inversion attacks have only been demonstrated on simple models, such as linear regression and logistic regression. Previous attempts to invert neural networks, even the ones with simple architectures, have failed to produce convincing results. Here we present a novel attack method, termed the \emph{generative model-inversion attack}, which can invert deep neural networks with high success rates. Rather than reconstructing private training data from scratch, we leverage partial public information, which can be very generic, to learn a distributional prior via generative adversarial networks (GANs) and use it to guide the inversion process. Moreover, we theoretically prove that a model's predictive power and its vulnerability to inversion attacks are indeed two sides of the same coin---highly predictive models are able to establish a strong correlation between features and labels, which coincides exactly with what an adversary exploits to mount the attacks. Our extensive experiments demonstrate that the proposed attack improves identification accuracy over the existing work by about $75\%$ for reconstructing face images from a state-of-the-art face recognition classifier. We also show that differential privacy, in its canonical form, is of little avail to defend against our attacks.