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.

The advent of text-to-image generative models has led artists to worry that their individual styles may be copied, creating a pressing need to reconsider the lack of protection for artistic styles under copyright law. This requires answering challenging questions, like what defines style and what constitutes style infringment. In this work, we build on prior legal scholarship to develop an automatic and interpretable framework to \emph{quantitatively} assess style infringement. Our methods hinge on a simple logical argument: if an artist's works can consistently be recognized as their own, then they have a unique style. Based on this argument, we introduce ArtSavant, a practical (i.e., efficient and easy to understand) tool to (i) determine the unique style of an artist by comparing it to a reference corpus of works from hundreds of artists, and (ii) recognize if the identified style reappears in generated images. We then apply ArtSavant in an empirical study to quantify the prevalence of artistic style copying across 3 popular text-to-image generative models, finding that under simple prompting, $20\\%$ of $372$ prolific artists studied appear to have their styles be at risk of copying by today's generative models. Our findings show that prior legal arguments can be operationalized in quantitative ways, towards more nuanced examination of the issue of artistic style infringements.

With models getting stronger, evaluations have grown more complex, testing multiple skills in one benchmark and even in the same instance at once. However, skill-wise performance is obscured when inspecting aggregate accuracy, under-utilizing the rich signal modern benchmarks contain. We propose an automatic approach to recover the underlying skills relevant for any evaluation instance, by way of inspecting model-generated {\em rationales}. After validating the relevance of rationale-parsed skills and inferring skills for $46$k instances over $12$ benchmarks, we observe many skills to be common across benchmarks, resulting in the curation of hundreds of \emph{skill-slices} (i.e. sets of instances testing a common skill). Inspecting accuracy over these slices yields novel insights on model trade-offs: e.g., compared to GPT-4o and Claude 3.5 Sonnet, on average, Gemini 1.5 Pro is $18\%$ more accurate in \emph{computing molar mass}, but $19\\%$ less accurate in \emph{applying constitutional law}, despite the overall accuracies of the three models differing by a mere $0.4\\%$. Furthermore, we demonstrate the practical utility of our approach by showing that insights derived from skill slice analysis can generalize to held-out instances: when routing each instance to the model strongest on the relevant skills, we see a $3\\%$ accuracy improvement over our $12$ dataset corpus. Our skill-slices and framework open a new avenue in model evaluation, leveraging skill-specific analyses to unlock a more granular and actionable understanding of model capabilities.

In this work, we study the challenge of providing human-understandable descriptions for failure modes in trained image classification models. Existing works address this problem by first identifying clusters (or directions) of incorrectly classified samples in a latent space and then aiming to provide human-understandable text descriptions for them. We observe that in some cases, describing text does not match well with identified failure modes, partially owing to the fact that shared interpretable attributes of failure modes may not be captured using clustering in the feature space. To improve on these shortcomings, we propose a novel approach that prioritizes interpretability in this problem: we start by obtaining human-understandable concepts (tags) of images in the dataset and then analyze the model's behavior based on the presence or absence of combinations of these tags. Our method also ensures that the tags describing a failure mode form a minimal set, avoiding redundant and noisy descriptions. Through several experiments on different datasets, we show that our method successfully identifies failure modes and generates high-quality text descriptions associated with them. These results highlight the importance of prioritizing interpretability in understanding model failures.

Deep neural networks can be unreliable in the real world when the training set does not adequately cover all the settings where they are deployed. Focusing on image classification, we consider the setting where we have an error distribution $\mathcal{E}$ representing a deployment scenario where the model fails. We have access to a small set of samples ${\mathcal{E}_{{\text{sample}}}}$ from $\mathcal{E}$ and it can be expensive to obtain additional samples. In the traditional model development framework, mitigating failures of the model in $\mathcal{E}$ can be challenging and is often done in an ad hoc manner. In this paper, we propose a general methodology for model debugging that can systemically improve model performance on $\mathcal{E}$ while maintaining its performance on the original test set. Our key assumption is that we have access to a large pool of weakly (noisily) labeled data $\mathcal{F}$. However, naively adding $\mathcal{F}$ to the training would hurt model performance due to the large extent of label noise. Our Data-Centric Debugging (DCD) framework carefully creates a debug-train set by selecting images from $\mathcal{F}$ that are visually similar to the images in ${\mathcal{E}_{{\text{sample}}}}$. To do this, we use the ℓ2distance in the feature space (penultimate layer activations) of various models including ResNet, Robust ResNet and DINO where we observe DINO ViTs are significantly better at discovering similar images compared to Resnets. Compared to the baselines that maintain model performance on the test set, we achieve significantly (+9.45%) improved results on the debug-heldout sets.

We observe that the mapping between an image's representation in one model to its representation in another can be learned surprisingly well with just a linear layer, even across diverse models. Building on this observation, we propose *text-to-concept*, where features from a fixed pretrained model are aligned linearly to the CLIP space, so that text embeddings from CLIP's text encoder become directly comparable to the aligned features. With text-to-concept, we convert fixed off-the-shelf vision encoders to surprisingly strong zero-shot classifiers for free, with accuracy at times even surpassing that of CLIP, despite being much smaller models and trained on a small fraction of the data compared to CLIP. We show other immediate use-cases of text-to-concept, like building concept bottleneck models with no concept supervision, diagnosing distribution shifts in terms of human concepts, and retrieving images satisfying a set of text-based constraints. Lastly, we demonstrate the feasibility of *concept-to-text*, where vectors in a model's feature space are decoded by first aligning to the CLIP before being fed to a GPT-based generative model. Our work suggests existing deep models, with presumably diverse architectures and training, represent input samples relatively similarly, and a two-way communication across model representation spaces and to humans (through language) is viable.

In recent years, concept-based approaches have emerged as some of the most promising explainability methods to help us interpret the decisions of Artificial Neural Networks (ANNs). These methods seek to discover intelligible visual ``concepts'' buried within the complex patterns of ANN activations in two key steps: (1) concept extraction followed by (2) importance estimation. While these two steps are shared across methods, they all differ in their specific implementations. Here, we introduce a unifying theoretical framework that recast the first step -- concept extraction problem -- as a special case of **dictionary learning**, and we formalize the second step -- concept importance estimation -- as a more general form of **attribution method**. This framework offers several advantages as it allows us: (i) to propose new evaluation metrics for comparing different concept extraction approaches; (ii) to leverage modern attribution methods and evaluation metrics to extend and systematically evaluate state-of-the-art concept-based approaches and importance estimation techniques; (iii) to derive theoretical guarantees regarding the optimality of such methods. We further leverage our framework to try to tackle a crucial question in explainability: how to *efficiently* identify clusters of data points that are classified based on a similar shared strategy. To illustrate these findings and to highlight the main strategies of a model, we introduce a visual representation called the strategic cluster graph. Finally, we present Lens, a dedicated website that offers a complete compilation of these visualizations for all classes of the ImageNet dataset.

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.

While datasets with single-label supervision have propelled rapid advances in image classification, additional annotations are necessary in order to quantitatively assess how models make predictions. To this end, for a subset of ImageNet samples, we collect segmentation masks for the entire object and 18 informative attributes. We call this dataset RIVAL10 (RIch Visual Attributes with Localization), consisting of roughly 26k instances over 10 classes. Using RIVAL10, we evaluate the sensitivity of a broad set of models to noise corruptions in foregrounds, backgrounds and attributes. In our analysis, we consider diverse state-of-the-art architectures (ResNets, Transformers) and training procedures (CLIP, SimCLR, DeiT, Adversarial Training). We find that, somewhat surprisingly, in ResNets, adversarial training makes models more sensitive to the background compared to foreground than standard training. Similarly, contrastively-trained models also have lower relative foreground sensitivity in both transformers and ResNets. Lastly, we observe intriguing adaptive abilities of transformers to increase relative foreground sensitivity as corruption level increases. Using saliency methods, we automatically discover spurious features that drive the background sensitivity of models and assess alignment of saliency maps with foregrounds. Finally, we quantitatively study the attribution problem for neural features by comparing feature saliency with ground-truth localization of semantic attributes.

Deep classifiers are known to rely on spurious features, leading to reduced generalization. The severity of this problem varies significantly by class. We identify $15$ classes in ImageNet with very strong spurious cues, and collect segmentation masks for these challenging objects to form \emph{Hard ImageNet}. Leveraging noise, saliency, and ablation based metrics, we demonstrate that models rely on spurious features in Hard ImageNet far more than in RIVAL10, an ImageNet analog to CIFAR10. We observe Hard ImageNet objects are less centered and occupy much less space in their images than RIVAL10 objects, leading to greater spurious feature reliance. Further, we use robust neural features to automatically rank our images based on the degree of spurious cues present. Comparing images with high and low rankings within a class reveals the exact spurious features models rely upon, and shows reduced performance when spurious features are absent. With Hard ImageNet's image rankings, object segmentations, and our extensive evaluation suite, the community can begin to address the problem of learning to detect challenging objects \emph{for the right reasons}, despite the presence of strong spurious cues.

Several existing works study either adversarial or natural distributional robustness of deep neural networks separately. In practice, however, models need to enjoy both types of robustness to ensure reliability. In this work, we bridge this gap and show that in fact, {\it explicit tradeoffs} exist between adversarial and natural distributional robustness. We first consider a simple linear regression setting on Gaussian data with disjoint sets of \emph{core} and \emph{spurious} features. In this setting, through theoretical and empirical analysis, we show that (i) adversarial training with $\ell_1$ and $\ell_2$ norms increases the model reliance on spurious features; (ii) For $\ell_\infty$ adversarial training, spurious reliance only occurs when the scale of the spurious features is larger than that of the core features; (iii) adversarial training can have {\it an unintended consequence} in reducing distributional robustness, specifically when spurious correlations are changed in the new test domain. Next, we present extensive empirical evidence, using a test suite of twenty adversarially trained models evaluated on five benchmark datasets (ObjectNet, RIVAL10, Salient ImageNet-1M, ImageNet-9, Waterbirds), that adversarially trained classifiers rely on backgrounds more than their standardly trained counterparts, validating our theoretical results. We also show that spurious correlations in training data (when preserved in the test domain) can {\it improve} adversarial robustness, revealing that previous claims that adversarial vulnerability is rooted in spurious correlations are incomplete.

Adversarial robustness of deep models is pivotal in ensuring safe deployment in real world settings, but most modern defenses have narrow scope and expensive costs. In this paper, we propose a self-supervised method to detect adversarial attacks and classify them to their respective threat models, based on a linear model operating on the embed-dings from a pretrained self-supervised encoder. We use a SimCLR encoder in our experiments, since we show the SimCLR embedding distance is a good proxy for human perceptibility, enabling it to encapsulate many threat models at once. We call our method SimCat since it uses SimCLR encoder to catch and categorize various types of adversarial attacks, including ℓpand non-ℓpevasion attacks, as well as data poisonings. The simple nature of a linear classifier makes our method efficient in both time and sample complexity. For example, on SVHN, using only five pairs of clean and adversarial examples computed with a PGD-ℓ∞attack, SimCat’s detection accuracy is over 85%. More-over, on ImageNet, using only 25 examples from each threat model, SimCat can classify eight different attack types such as PGD-ℓ2, PGD-ℓ∞, CW-ℓ2, PPGD, LPA, StAdv, ReColor, and JPEG-ℓ∞, with over 40% accuracy. On STL10 data, we apply SimCat as a defense against poisoning attacks, such as BP, CP, FC, CLBD, HTBD, halving the success rate while using only twenty total poisons for training. We find that the detectors generalize well to unseen threat models. Lastly, we investigate the performance of our detection method under adaptive attacks and further boost its robustness against such attacks via adversarial training.