The integration of large language models (LLMs) into information retrieval systems introduces new attack surfaces, particularly for adversarial ranking manipulations. We present StealthRank, a novel adversarial ranking attack that manipulates LLM-driven product recommendation systems while maintaining textual fluency and stealth. Unlike existing methods that often introduce detectable anomalies, StealthRank employs an energy-based optimization framework combined with Langevin dynamics to generate StealthRank Prompts (SRPs)-adversarial text sequences embedded within product descriptions that subtly yet effectively influence LLM ranking mechanisms. We evaluate StealthRank across multiple LLMs, demonstrating its ability to covertly boost the ranking of target products while avoiding explicit manipulation traces that can be easily detected. Our results show that StealthRank consistently outperforms state-of-the-art adversarial ranking baselines in both effectiveness and stealth, highlighting critical vulnerabilities in LLM-driven recommendation systems.
In this paper, we systematically evaluate the robustness of multi-exit language models against adversarial slowdown. To audit their robustness, we design a slowdown attack that generates natural adversarial text bypassing early-exit points. We use the resulting WAFFLE attack as a vehicle to conduct a comprehensive evaluation of three multi-exit mechanisms with the GLUE benchmark against adversarial slowdown. We then show our attack significantly reduces the computational savings provided by the three methods in both white-box and black-box settings. The more complex a mechanism is, the more vulnerable it is to adversarial slowdown. We also perform a linguistic analysis of the perturbed text inputs, identifying common perturbation patterns that our attack generates, and comparing them with standard adversarial text attacks. Moreover, we show that adversarial training is ineffective in defeating our slowdown attack, but input sanitization with a conversational model, e.g., ChatGPT, can remove perturbations effectively. This result suggests that future work is needed for developing efficient yet robust multi-exit models. Our code is available at: https://github.com/ztcoalson/WAFFLE
Social media platforms like Twitter have increasingly relied on Natural Language Processing NLP techniques to analyze and understand the sentiments expressed in the user generated content. One such state of the art NLP model is Bidirectional Encoder Representations from Transformers BERT which has been widely adapted in sentiment analysis. BERT is susceptible to adversarial attacks. This paper aims to scrutinize the inherent vulnerabilities of such models in Twitter sentiment analysis. It aims to formulate a framework for constructing targeted adversarial texts capable of deceiving these models, while maintaining stealth. In contrast to conventional methodologies, such as Importance Reweighting, this framework core idea resides in its reliance on gradients to prioritize the importance of individual words within the text. It uses a whitebox approach to attain fine grained sensitivity, pinpointing words that exert maximal influence on the classification outcome. This paper is organized into three interdependent phases. It starts with fine-tuning a pre-trained BERT model on Twitter data. It then analyzes gradients of the model to rank words on their importance, and iteratively replaces those with feasible candidates until an acceptable solution is found. Finally, it evaluates the effectiveness of the adversarial text against the custom trained sentiment classification model. This assessment would help in gauging the capacity of the adversarial text to successfully subvert classification without raising any alarm.
To understand the complexity of sequence classification tasks, Hahn et al. (2021) proposed sensitivity as the number of disjoint subsets of the input sequence that can each be individually changed to change the output. Though effective, calculating sensitivity at scale using this framework is costly because of exponential time complexity. Therefore, we introduce a Sensitivity-based Multi-Armed Bandit framework (SMAB), which provides a scalable approach for calculating word-level local (sentence-level) and global (aggregated) sensitivities concerning an underlying text classifier for any dataset. We establish the effectiveness of our approach through various applications. We perform a case study on CHECKLIST generated sentiment analysis dataset where we show that our algorithm indeed captures intuitively high and low-sensitive words. Through experiments on multiple tasks and languages, we show that sensitivity can serve as a proxy for accuracy in the absence of gold data. Lastly, we show that guiding perturbation prompts using sensitivity values in adversarial example generation improves attack success rate by 15.58%, whereas using sensitivity as an additional reward in adversarial paraphrase generation gives a 12.00% improvement over SOTA approaches. Warning: Contains potentially offensive content.
Recent research has shown that neural information retrieval techniques may be susceptible to adversarial attacks. Adversarial attacks seek to manipulate the ranking of documents, with the intention of exposing users to targeted content. In this paper, we introduce the Embedding Perturbation Rank Attack (EMPRA) method, a novel approach designed to perform adversarial attacks on black-box Neural Ranking Models (NRMs). EMPRA manipulates sentence-level embeddings, guiding them towards pertinent context related to the query while preserving semantic integrity. This process generates adversarial texts that seamlessly integrate with the original content and remain imperceptible to humans. Our extensive evaluation conducted on the widely-used MS MARCO V1 passage collection demonstrate the effectiveness of EMPRA against a wide range of state-of-the-art baselines in promoting a specific set of target documents within a given ranked results. Specifically, EMPRA successfully achieves a re-ranking of almost 96% of target documents originally ranked between 51-100 to rank within the top 10. Furthermore, EMPRA does not depend on surrogate models for adversarial text generation, enhancing its robustness against different NRMs in realistic settings.
Adversarial text attack research is useful for testing the robustness of NLP models, however, the rise of transformers has greatly increased the time required to test attacks. Especially when researchers do not have access to adequate resources (e.g. GPUs). This can hinder attack research, as modifying one example for an attack can require hundreds of queries to a model, especially for black-box attacks. Often these attacks remove one token at a time to find the ideal one to change, requiring $n$ queries (the length of the text) right away. We propose a more efficient selection method called BinarySelect which combines binary search and attack selection methods to greatly reduce the number of queries needed to find a token. We find that BinarySelect only needs $\text{log}_2(n) * 2$ queries to find the first token compared to $n$ queries. We also test BinarySelect in an attack setting against 5 classifiers across 3 datasets and find a viable tradeoff between number of queries saved and attack effectiveness. For example, on the Yelp dataset, the number of queries is reduced by 32% (72 less) with a drop in attack effectiveness of only 5 points. We believe that BinarySelect can help future researchers study adversarial attacks and black-box problems more efficiently and opens the door for researchers with access to less resources.
Language models based on deep neural networks are vulnerable to textual adversarial attacks. While rich-resource languages like English are receiving focused attention, Tibetan, a cross-border language, is gradually being studied due to its abundant ancient literature and critical language strategy. Currently, there are several Tibetan adversarial text generation methods, but they do not fully consider the textual features of Tibetan script and overestimate the quality of generated adversarial texts. To address this issue, we propose a novel Tibetan adversarial text generation method called TSCheater, which considers the characteristic of Tibetan encoding and the feature that visually similar syllables have similar semantics. This method can also be transferred to other abugidas, such as Devanagari script. We utilize a self-constructed Tibetan syllable visual similarity database called TSVSDB to generate substitution candidates and adopt a greedy algorithm-based scoring mechanism to determine substitution order. After that, we conduct the method on eight victim language models. Experimentally, TSCheater outperforms existing methods in attack effectiveness, perturbation magnitude, semantic similarity, visual similarity, and human acceptance. Finally, we construct the first Tibetan adversarial robustness evaluation benchmark called AdvTS, which is generated by existing methods and proofread by humans.
Neural Machine Translation systems are used in diverse applications due to their impressive performance. However, recent studies have shown that these systems are vulnerable to carefully crafted small perturbations to their inputs, known as adversarial attacks. In this paper, we propose a new type of adversarial attack against NMT models. In this attack, we find a word to be added between two sentences such that the second sentence is ignored and not translated by the NMT model. The word added between the two sentences is such that the whole adversarial text is natural in the source language. This type of attack can be harmful in practical scenarios since the attacker can hide malicious information in the automatic translation made by the target NMT model. Our experiments show that different NMT models and translation tasks are vulnerable to this type of attack. Our attack can successfully force the NMT models to ignore the second part of the input in the translation for more than 50% of all cases while being able to maintain low perplexity for the whole input.
Recent advances in diffusion models have significantly enhanced the quality of image synthesis, yet they have also introduced serious safety concerns, particularly the generation of Not Safe for Work (NSFW) content. Previous research has demonstrated that adversarial prompts can be used to generate NSFW content. However, such adversarial text prompts are often easily detectable by text-based filters, limiting their efficacy. In this paper, we expose a previously overlooked vulnerability: adversarial image attacks targeting Image-to-Image (I2I) diffusion models. We propose AdvI2I, a novel framework that manipulates input images to induce diffusion models to generate NSFW content. By optimizing a generator to craft adversarial images, AdvI2I circumvents existing defense mechanisms, such as Safe Latent Diffusion (SLD), without altering the text prompts. Furthermore, we introduce AdvI2I-Adaptive, an enhanced version that adapts to potential countermeasures and minimizes the resemblance between adversarial images and NSFW concept embeddings, making the attack more resilient against defenses. Through extensive experiments, we demonstrate that both AdvI2I and AdvI2I-Adaptive can effectively bypass current safeguards, highlighting the urgent need for stronger security measures to address the misuse of I2I diffusion models.
We design, implement, and evaluate adversarial decoding, a new, generic text generation technique that produces readable documents for different adversarial objectives. Prior methods either produce easily detectable gibberish, or cannot handle objectives that include embedding similarity. In particular, they only work for direct attacks (such as jailbreaking) and cannot produce adversarial text for realistic indirect injection, e.g., documents that (1) are retrieved in RAG systems in response to broad classes of queries, and also (2) adversarially influence subsequent generation. We also show that fluency (low perplexity) is not sufficient to evade filtering. We measure the effectiveness of adversarial decoding for different objectives, including RAG poisoning, jailbreaking, and evasion of defensive filters, and demonstrate that it outperforms existing methods while producing readable adversarial documents.
While neural machine translation (NMT) models achieve success in our daily lives, they show vulnerability to adversarial attacks. Despite being harmful, these attacks also offer benefits for interpreting and enhancing NMT models, thus drawing increased research attention. However, existing studies on adversarial attacks are insufficient in both attacking ability and human imperceptibility due to their sole focus on the scope of language. This paper proposes a novel vision-fused attack (VFA) framework to acquire powerful adversarial text, i.e., more aggressive and stealthy. Regarding the attacking ability, we design the vision-merged solution space enhancement strategy to enlarge the limited semantic solution space, which enables us to search for adversarial candidates with higher attacking ability. For human imperceptibility, we propose the perception-retained adversarial text selection strategy to align the human text-reading mechanism. Thus, the finally selected adversarial text could be more deceptive. Extensive experiments on various models, including large language models (LLMs) like LLaMA and GPT-3.5, strongly support that VFA outperforms the comparisons by large margins (up to 81%/14% improvements on ASR/SSIM).
Text-aware recommender systems incorporate rich textual features, such as titles and descriptions, to generate item recommendations for users. The use of textual features helps mitigate cold-start problems, and thus, such recommender systems have attracted increased attention. However, we argue that the dependency on item descriptions makes the recommender system vulnerable to manipulation by adversarial sellers on e-commerce platforms. In this paper, we explore the possibility of such manipulation by proposing a new text rewriting framework to attack text-aware recommender systems. We show that the rewriting attack can be exploited by sellers to unfairly uprank their products, even though the adversarially rewritten descriptions are perceived as realistic by human evaluators. Methodologically, we investigate two different variations to carry out text rewriting attacks: (1) two-phase fine-tuning for greater attack performance, and (2) in-context learning for higher text rewriting quality. Experiments spanning 3 different datasets and 4 existing approaches demonstrate that recommender systems exhibit vulnerability against the proposed text rewriting attack. Our work adds to the existing literature around the robustness of recommender systems, while highlighting a new dimension of vulnerability in the age of large-scale automated text generation.
As language models (LMs) are used to build autonomous agents in real environments, ensuring their adversarial robustness becomes a critical challenge. Unlike chatbots, agents are compound systems with multiple components taking actions, which existing LMs safety evaluations do not adequately address. To bridge this gap, we manually create 200 targeted adversarial tasks and evaluation scripts in a realistic threat model on top of VisualWebArena, a real environment for web agents. To systematically examine the robustness of agents, we propose the Agent Robustness Evaluation (ARE) framework. ARE views the agent as a graph showing the flow of intermediate outputs between components and decomposes robustness as the flow of adversarial information on the graph. We find that we can successfully break latest agents that use black-box frontier LMs, including those that perform reflection and tree search. With imperceptible perturbations to a single image (less than 5% of total web page pixels), an attacker can hijack these agents to execute targeted adversarial goals with success rates up to 67%. We also use ARE to rigorously evaluate how the robustness changes as new components are added. We find that inference-time compute that typically improves benign performance can open up new vulnerabilities and harm robustness. An attacker can compromise the evaluator used by the reflexion agent and the value function of the tree search agent, which increases the attack success relatively by 15% and 20%. Our data and code for attacks, defenses, and evaluation are at https://github.com/ChenWu98/agent-attack
Recent advancements in Large Vision-Language Models (VLMs) have underscored their superiority in various multimodal tasks. However, the adversarial robustness of VLMs has not been fully explored. Existing methods mainly assess robustness through unimodal adversarial attacks that perturb images, while assuming inherent resilience against text-based attacks. Different from existing attacks, in this work we propose a more comprehensive strategy that jointly attacks both text and image modalities to exploit a broader spectrum of vulnerability within VLMs. Specifically, we propose a dual optimization objective aimed at guiding the model to generate affirmative responses with high toxicity. Our attack method begins by optimizing an adversarial image prefix from random noise to generate diverse harmful responses in the absence of text input, thus imbuing the image with toxic semantics. Subsequently, an adversarial text suffix is integrated and co-optimized with the adversarial image prefix to maximize the probability of eliciting affirmative responses to various harmful instructions. The discovered adversarial image prefix and text suffix are collectively denoted as a Universal Master Key (UMK). When integrated into various malicious queries, UMK can circumvent the alignment defenses of VLMs and lead to the generation of objectionable content, known as jailbreaks. The experimental results demonstrate that our universal attack strategy can effectively jailbreak MiniGPT-4 with a 96% success rate, highlighting the vulnerability of VLMs and the urgent need for new alignment strategies.
Pretrained vision-language models (VLMs) like CLIP exhibit exceptional generalization across diverse downstream tasks. While recent studies reveal their vulnerability to adversarial attacks, research to date has primarily focused on enhancing the robustness of image encoders against image-based attacks, with defenses against text-based and multimodal attacks remaining largely unexplored. To this end, this work presents the first comprehensive study on improving the adversarial robustness of VLMs against attacks targeting image, text, and multimodal inputs. This is achieved by proposing multimodal contrastive adversarial training (MMCoA). Such an approach strengthens the robustness of both image and text encoders by aligning the clean text embeddings with adversarial image embeddings, and adversarial text embeddings with clean image embeddings. The robustness of the proposed MMCoA is examined against existing defense methods over image, text, and multimodal attacks on the CLIP model. Extensive experiments on 15 datasets across two tasks reveal the characteristics of different adversarial defense methods under distinct distribution shifts and dataset complexities across the three attack types. This paves the way for a unified framework of adversarial robustness against different modality attacks, opening up new possibilities for securing VLMs against multimodal attacks. The code is available at https://github.com/ElleZWQ/MMCoA.git.
The vulnerability of deep neural networks to imperceptible adversarial perturbations has attracted widespread attention. Inspired by the success of vision-language foundation models, previous efforts achieved zero-shot adversarial robustness by aligning adversarial visual features with text supervision. However, in practice, they are still unsatisfactory due to several issues, including heavy adaptation cost, suboptimal text supervision, and uncontrolled natural generalization capacity. In this paper, to address these issues, we propose a few-shot adversarial prompt framework where adapting input sequences with limited data makes significant adversarial robustness improvement. Specifically, we achieve this by providing adversarially correlated text supervision that is end-to-end learned from adversarial examples. We also propose a novel training objective that enhances the consistency of multi-modal features while encourages differentiated uni-modal features between natural and adversarial examples. The proposed framework gives access to learn adversarial text supervision, which provides superior cross-modal adversarial alignment and matches state-of-the-art zero-shot adversarial robustness with only 1% training data. Code is available at: https://github.com/lionel-w2/FAP.
Vision-language pre-training (VLP) models exhibit remarkable capabilities in comprehending both images and text, yet they remain susceptible to multimodal adversarial examples (AEs). Strengthening attacks and uncovering vulnerabilities, especially common issues in VLP models (e.g., high transferable AEs), can advance reliable and practical VLP models. A recent work (i.e., Set-level guidance attack) indicates that augmenting image-text pairs to increase AE diversity along the optimization path enhances the transferability of adversarial examples significantly. However, this approach predominantly emphasizes diversity around the online adversarial examples (i.e., AEs in the optimization period), leading to the risk of overfitting the victim model and affecting the transferability. In this study, we posit that the diversity of adversarial examples towards the clean input and online AEs are both pivotal for enhancing transferability across VLP models. Consequently, we propose using diversification along the intersection region of adversarial trajectory to expand the diversity of AEs. To fully leverage the interaction between modalities, we introduce text-guided adversarial example selection during optimization. Furthermore, to further mitigate the potential overfitting, we direct the adversarial text deviating from the last intersection region along the optimization path, rather than adversarial images as in existing methods. Extensive experiments affirm the effectiveness of our method in improving transferability across various VLP models and downstream vision-and-language tasks.
Existing works have shown that fine-tuned textual transformer models achieve state-of-the-art prediction performances but are also vulnerable to adversarial text perturbations. Traditional adversarial evaluation is often done \textit{only after} fine-tuning the models and ignoring the training data. In this paper, we want to prove that there is also a strong correlation between training data and model robustness. To this end, we extract 13 different features representing a wide range of input fine-tuning corpora properties and use them to predict the adversarial robustness of the fine-tuned models. Focusing mostly on encoder-only transformer models BERT and RoBERTa with additional results for BART, ELECTRA, and GPT2, we provide diverse evidence to support our argument. First, empirical analyses show that (a) extracted features can be used with a lightweight classifier such as Random Forest to predict the attack success rate effectively, and (b) features with the most influence on the model robustness have a clear correlation with the robustness. Second, our framework can be used as a fast and effective additional tool for robustness evaluation since it (a) saves 30x-193x runtime compared to the traditional technique, (b) is transferable across models, (c) can be used under adversarial training, and (d) robust to statistical randomness. Our code is publicly available at \url{https://github.com/CaptainCuong/RobustText_ACL2024}.
Text classification systems have been proven vulnerable to adversarial text examples, modified versions of the original text examples that are often unnoticed by human eyes, yet can force text classification models to alter their classification. Often, research works quantifying the impact of adversarial text attacks have been applied only to models trained in English. In this paper, we introduce the first word-level study of adversarial attacks in Arabic. Specifically, we use a synonym (word-level) attack using a Masked Language Modeling (MLM) task with a BERT model in a black-box setting to assess the robustness of the state-of-the-art text classification models to adversarial attacks in Arabic. To evaluate the grammatical and semantic similarities of the newly produced adversarial examples using our synonym BERT-based attack, we invite four human evaluators to assess and compare the produced adversarial examples with their original examples. We also study the transferability of these newly produced Arabic adversarial examples to various models and investigate the effectiveness of defense mechanisms against these adversarial examples on the BERT models. We find that fine-tuned BERT models were more susceptible to our synonym attacks than the other Deep Neural Networks (DNN) models like WordCNN and WordLSTM we trained. We also find that fine-tuned BERT models were more susceptible to transferred attacks. We, lastly, find that fine-tuned BERT models successfully regain at least 2% in accuracy after applying adversarial training as an initial defense mechanism.
This paper describes RETVec, an efficient, resilient, and multilingual text vectorizer designed for neural-based text processing. RETVec combines a novel character encoding with an optional small embedding model to embed words into a 256-dimensional vector space. The RETVec embedding model is pre-trained using pair-wise metric learning to be robust against typos and character-level adversarial attacks. In this paper, we evaluate and compare RETVec to state-of-the-art vectorizers and word embeddings on popular model architectures and datasets. These comparisons demonstrate that RETVec leads to competitive, multilingual models that are significantly more resilient to typos and adversarial text attacks. RETVec is available under the Apache 2 license at https://github.com/google-research/retvec.