The first is a traditional Label Attack, tricking VLMs into misidentifying class labels, such as confusing Donald Trump for Joe Biden.

Vision-Language Models (VLMs) excel in generating textual responses from visual inputs, but their versatility raises security concerns. This study takes the first step in exposing VLMs' susceptibility to data poisoning attacks that can manipulate responses to innocuous, everyday prompts. We introduce Shadowcast, a stealthy data poisoning attack where poison samples are visually indistinguishable from benign images with matching texts. Shadowcast demonstrates effectiveness in two attack types. The first is a traditional Label Attack, tricking VLMs into misidentifying class labels, such as confusing Donald Trump for Joe Biden.

Poisoning attacks can compromise the safety of large language models (LLMs) by injecting malicious documents into their training data. Existing work has studied pretraining poisoning assuming adversaries control a percentage of the training corpus. However, for large models, even small percentages translate to impractically large amounts of data. This work demonstrates for the first time that poisoning attacks instead require a near-constant number of documents regardless of dataset size. We conduct the largest pretraining poisoning experiments to date, pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes, despite the largest models training on more than 20 times more clean data. We also run smaller-scale experiments to ablate factors that could influence attack success, including broader ratios of poisoned to clean data and non-random distributions of poisoned samples. Finally, we demonstrate the same dynamics for poisoning during fine-tuning. Altogether, our results suggest that injecting backdoors through data poisoning may be easier for large models than previously believed as the number of poisons required does not scale up with model size, highlighting the need for more research on defences to mitigate this risk in future models.

Today's text-to-image generative models are trained on millions of images sourced from the Internet, each paired with a detailed caption produced by Vision-Language Models (VLMs). This part of the training pipeline is critical for supplying the models with large volumes of high-quality image-caption pairs during training. However, recent work suggests that VLMs are vulnerable to stealthy adversarial attacks, where adversarial perturbations are added to images to mislead the VLMs into producing incorrect captions. In this paper, we explore the feasibility of adversarial mislabeling attacks on VLMs as a mechanism to poisoning training pipelines for text-to-image models. Our experiments demonstrate that VLMs are highly vulnerable to adversarial perturbations, allowing attackers to produce benign-looking images that are consistently miscaptioned by the VLM models. This has the effect of injecting strong "dirty-label" poison samples into the training pipeline for text-to-image models, successfully altering their behavior with a small number of poisoned samples. We find that while potential defenses can be effective, they can be targeted and circumvented by adaptive attackers. This suggests a cat-and-mouse game that is likely to reduce the quality of training data and increase the cost of text-to-image model development. Finally, we demonstrate the real-world effectiveness of these attacks, achieving high attack success (over 73%) even in black-box scenarios against commercial VLMs (Google Vertex AI and Microsoft Azure).

Vision-Language Models (VLMs) excel in generating textual responses from visual inputs, but their versatility raises security concerns. This study takes the first step in exposing VLMs' susceptibility to data poisoning attacks that can manipulate responses to innocuous, everyday prompts. We introduce Shadowcast, a stealthy data poisoning attack where poison samples are visually indistinguishable from benign images with matching texts. Shadowcast demonstrates effectiveness in two attack types. The first is a traditional Label Attack, tricking VLMs into misidentifying class labels, such as confusing Donald Trump for Joe Biden.

Backdoor attacks on text classifiers can cause them to predict a predefined label when a particular "trigger" is present. Prior attacks often rely on triggers that are ungrammatical or otherwise unusual, leading to conspicuous attacks. As a result, human annotators, who play a critical role in curating training data in practice, can easily detect and filter out these unnatural texts during manual inspection, reducing the risk of such attacks. We argue that a key criterion for a successful attack is for text with and without triggers to be indistinguishable to humans. However, prior work neither directly nor comprehensively evaluated attack subtlety and invisibility with human involvement. We bridge the gap by conducting thorough human evaluations to assess attack subtlety. We also propose \emph{AttrBkd}, consisting of three recipes for crafting subtle yet effective trigger attributes, such as extracting fine-grained attributes from existing baseline backdoor attacks. Our human evaluations find that AttrBkd with these baseline-derived attributes is often more effective (higher attack success rate) and more subtle (fewer instances detected by humans) than the original baseline backdoor attacks, demonstrating that backdoor attacks can bypass detection by being inconspicuous and appearing natural even upon close inspection, while still remaining effective. Our human annotation also provides information not captured by automated metrics used in prior work, and demonstrates the misalignment of these metrics with human judgment.

As the capabilities of large language models (LLMs) continue to expand, their usage has become increasingly prevalent. However, as reflected in numerous ongoing lawsuits regarding LLM-generated content, addressing copyright infringement remains a significant challenge. In this paper, we introduce PoisonedParrot: the first stealthy data poisoning attack that induces an LLM to generate copyrighted content even when the model has not been directly trained on the specific copyrighted material. PoisonedParrot integrates small fragments of copyrighted text into the poison samples using an off-the-shelf LLM. Despite its simplicity, evaluated in a wide range of experiments, PoisonedParrot is surprisingly effective at priming the model to generate copyrighted content with no discernible side effects. Moreover, we discover that existing defenses are largely ineffective against our attack. Finally, we make the first attempt at mitigating copyright-infringement poisoning attacks by proposing a defense: ParrotTrap. We encourage the community to explore this emerging threat model further.

Backdoor attacks embed hidden functionalities in deep neural networks (DNN), triggering malicious behavior with specific inputs. Advanced defenses monitor anomalous DNN inferences to detect such attacks. However, concealed backdoors evade detection by maintaining a low pre-deployment attack success rate (ASR) and restoring high ASR post-deployment via machine unlearning. Existing concealed backdoors are often constrained by requiring white-box or black-box access or auxiliary data, limiting their practicality when such access or data is unavailable. This paper introduces ReVeil, a concealed backdoor attack targeting the data collection phase of the DNN training pipeline, requiring no model access or auxiliary data. ReVeil maintains low pre-deployment ASR across four datasets and four trigger patterns, successfully evades three popular backdoor detection methods, and restores high ASR post-deployment through machine unlearning.

Targeted poisoning attacks aim to compromise the model's prediction on specific target samples. In a common clean-label setting, they are achieved by slightly perturbing a subset of training samples given access to those specific targets. Despite continuous efforts, it remains unexplored whether such attacks can generalize to unknown variations of those targets. In this paper, we take the first step to systematically study this generalization problem. Observing that the widely adopted, cosine similarity-based attack exhibits limited generalizability, we propose a well-generalizable attack that leverages both the direction and magnitude of model gradients. In particular, we explore diverse target variations, such as an object with varied viewpoints and an animal species with distinct appearances. Extensive experiments across various generalization scenarios demonstrate that our method consistently achieves the best attack effectiveness. For example, our method outperforms the cosine similarity-based attack by 20.95% in attack success rate with similar overall accuracy, averaged over four models on two image benchmark datasets. The code is available at https://github.com/jiaangk/generalizable_tcpa