Chejian Xu

Boxin Wang, Weixin Chen, Hengzhi Pei et al. · berkeley, microsoft-research

Generative Pre-trained Transformer (GPT) models have exhibited exciting progress in their capabilities, capturing the interest of practitioners and the public alike. Yet, while the literature on the trustworthiness of GPT models remains limited, practitioners have proposed employing capable GPT models for sensitive applications such as healthcare and finance -- where mistakes can be costly. To this end, this work proposes a comprehensive trustworthiness evaluation for large language models with a focus on GPT-4 and GPT-3.5, considering diverse perspectives -- including toxicity, stereotype bias, adversarial robustness, out-of-distribution robustness, robustness on adversarial demonstrations, privacy, machine ethics, and fairness. Based on our evaluations, we discover previously unpublished vulnerabilities to trustworthiness threats. For instance, we find that GPT models can be easily misled to generate toxic and biased outputs and leak private information in both training data and conversation history. We also find that although GPT-4 is usually more trustworthy than GPT-3.5 on standard benchmarks, GPT-4 is more vulnerable given jailbreaking system or user prompts, potentially because GPT-4 follows (misleading) instructions more precisely. Our work illustrates a comprehensive trustworthiness evaluation of GPT models and sheds light on the trustworthiness gaps. Our benchmark is publicly available at https://decodingtrust.github.io/ ; our dataset can be previewed at https://huggingface.co/datasets/AI-Secure/DecodingTrust ; a concise version of this work is at https://openreview.net/pdf?id=kaHpo8OZw2 .

Peng Xu, Wei Ping, Xianchao Wu et al.

In this work, we introduce ChatQA 2, an Llama 3.0-based model with a 128K context window, designed to bridge the gap between open-source LLMs and leading proprietary models (e.g., GPT-4-Turbo-2024-04-09) in long context understanding and retrieval-augmented generation (RAG) capabilities. These two capabilities are complementary to each other and essential for LLMs to process large volumes of information that cannot fit into a single prompt. We present a detailed continued training recipe to extend the context window of Llama3-70B-base from 8K to 128K tokens, along with a three-stage instruction tuning process to enhance the model's instruction-following, RAG performance, and long-context understanding capabilities. Our results demonstrate that the Llama3-ChatQA-2-70B model outperforms most existing state-of-the-art models, including GPT-4-Turbo-2024-04-09, Qwen2-72B-Instruct, and Llama3.1-70B-Instruct, on ultra-long tasks beyond 100K tokens, as well as on the RAG benchmark using only a 4K context window, showing the strong long context capability across varying sequence lengths. We further provide extensive comparisons between direct long-context and RAG solutions using the same state-of-the-art long-context LLMs. Interestingly, we find that the performance of strong long-context LLMs using RAG improves when retrieving a larger number of chunks. With a large set of top-k chunks, RAG consistently outperforms direct long-context solution using the same state-of-the-art long-context models (e.g., Llama3-ChatQA-2-70B and Qwen2-72B-Instruct) on both 32K and 128K benchmarks. We open-source the model weights, training data, and the evaluation setup for the for the community: https://chatqa2-project.github.io/

Boxin Wang, Chejian Xu, Xiangyu Liu et al.

Recent studies show that pre-trained language models (LMs) are vulnerable to textual adversarial attacks. However, existing attack methods either suffer from low attack success rates or fail to search efficiently in the exponentially large perturbation space. We propose an efficient and effective framework SemAttack to generate natural adversarial text by constructing different semantic perturbation functions. In particular, SemAttack optimizes the generated perturbations constrained on generic semantic spaces, including typo space, knowledge space (e.g., WordNet), contextualized semantic space (e.g., the embedding space of BERT clusterings), or the combination of these spaces. Thus, the generated adversarial texts are more semantically close to the original inputs. Extensive experiments reveal that state-of-the-art (SOTA) large-scale LMs (e.g., DeBERTa-v2) and defense strategies (e.g., FreeLB) are still vulnerable to SemAttack. We further demonstrate that SemAttack is general and able to generate natural adversarial texts for different languages (e.g., English and Chinese) with high attack success rates. Human evaluations also confirm that our generated adversarial texts are natural and barely affect human performance. Our code is publicly available at https://github.com/AI-secure/SemAttack.

Zhenguang Liu, Sifan Wu, Chejian Xu et al.

One compelling application of artificial intelligence is to generate a video of a target person performing arbitrary desired motion (from a source person). While the state-of-the-art methods are able to synthesize a video demonstrating similar broad stroke motion details, they are generally lacking in texture details. A pertinent manifestation appears as distorted face, feet, and hands, and such flaws are very sensitively perceived by human observers. Furthermore, current methods typically employ GANs with a L2 loss to assess the authenticity of the generated videos, inherently requiring a large amount of training samples to learn the texture details for adequate video generation. In this work, we tackle these challenges from three aspects: 1) We disentangle each video frame into foreground (the person) and background, focusing on generating the foreground to reduce the underlying dimension of the network output. 2) We propose a theoretically motivated Gromov-Wasserstein loss that facilitates learning the mapping from a pose to a foreground image. 3) To enhance texture details, we encode facial features with geometric guidance and employ local GANs to refine the face, feet, and hands. Extensive experiments show that our method is able to generate realistic target person videos, faithfully copying complex motions from a source person.

Fan Wu, Linyi Li, Chejian Xu et al.

As reinforcement learning (RL) has achieved near human-level performance in a variety of tasks, its robustness has raised great attention. While a vast body of research has explored test-time (evasion) attacks in RL and corresponding defenses, its robustness against training-time (poisoning) attacks remains largely unanswered. In this work, we focus on certifying the robustness of offline RL in the presence of poisoning attacks, where a subset of training trajectories could be arbitrarily manipulated. We propose the first certification framework, COPA, to certify the number of poisoning trajectories that can be tolerated regarding different certification criteria. Given the complex structure of RL, we propose two certification criteria: per-state action stability and cumulative reward bound. To further improve the certification, we propose new partition and aggregation protocols to train robust policies. We further prove that some of the proposed certification methods are theoretically tight and some are NP-Complete problems. We leverage COPA to certify three RL environments trained with different algorithms and conclude: (1) The proposed robust aggregation protocols such as temporal aggregation can significantly improve the certifications; (2) Our certification for both per-state action stability and cumulative reward bound are efficient and tight; (3) The certification for different training algorithms and environments are different, implying their intrinsic robustness properties. All experimental results are available at https://copa-leaderboard.github.io.

Zhaorun Chen, Xun Liu, Haibo Tong et al.

AI agents are increasingly deployed across diverse domains to automate complex workflows through long-horizon and high-stakes action executions. Due to their high capability and flexibility, such agents raise significant security and safety concerns. A growing number of real-world incidents have shown that adversaries can easily manipulate agents into performing harmful actions, such as leaking API keys, deleting user data, or initiating unauthorized transactions. Evaluating agent security is inherently challenging, as agents operate in dynamic, untrusted environments involving external tools, heterogeneous data sources, and frequent user interactions. However, realistic, controllable, and reproducible environments for large-scale risk assessment remain largely underexplored. To address this gap, we introduce the DecodingTrust-Agent Platform (DTap), the first controllable and interactive red-teaming platform for AI agents, spanning 14 real-world domains and over 50 simulation environments that replicate widely used systems such as Google Workspace, Paypal, and Slack. To scale the risk assessment of agents in DTap, we further propose DTap-Red, the first autonomous red-teaming agent that systematically explores diverse injection vectors (e.g., prompt, tool, skill, environment, combinations) and autonomously discovers effective attack strategies tailored to varying malicious goals. Using DTap-Red, we curate DTap-Bench, a large-scale red-teaming dataset comprising high-quality instances across domains, each paired with a verifiable judge to automatically validate attack outcomes. Through DTap, we conduct large-scale evaluations of popular AI agents built on various backbone models, spanning security policies, risk categories, and attack strategies, revealing systematic vulnerability patterns and providing valuable insights for developing secure next-generation agents.

Zeyi Liao, Lingbo Mo, Chejian Xu et al.

Generalist web agents have demonstrated remarkable potential in autonomously completing a wide range of tasks on real websites, significantly boosting human productivity. However, web tasks, such as booking flights, usually involve users' PII, which may be exposed to potential privacy risks if web agents accidentally interact with compromised websites, a scenario that remains largely unexplored in the literature. In this work, we narrow this gap by conducting the first study on the privacy risks of generalist web agents in adversarial environments. First, we present a realistic threat model for attacks on the website, where we consider two adversarial targets: stealing users' specific PII or the entire user request. Then, we propose a novel attack method, termed Environmental Injection Attack (EIA). EIA injects malicious content designed to adapt well to environments where the agents operate and our work instantiates EIA specifically for privacy scenarios in web environments. We collect 177 action steps that involve diverse PII categories on realistic websites from the Mind2Web, and conduct experiments using one of the most capable generalist web agent frameworks to date. The results demonstrate that EIA achieves up to 70% ASR in stealing specific PII and 16% ASR for full user request. Additionally, by accessing the stealthiness and experimenting with a defensive system prompt, we indicate that EIA is hard to detect and mitigate. Notably, attacks that are not well adapted for a webpage can be detected via human inspection, leading to our discussion about the trade-off between security and autonomy. However, extra attackers' efforts can make EIA seamlessly adapted, rendering such supervision ineffective. Thus, we further discuss the defenses at the pre- and post-deployment stages of the websites without relying on human supervision and call for more advanced defense strategies.

Jiawei Zhang, Chejian Xu, Bo Li

We present ChatScene, a Large Language Model (LLM)-based agent that leverages the capabilities of LLMs to generate safety-critical scenarios for autonomous vehicles. Given unstructured language instructions, the agent first generates textually described traffic scenarios using LLMs. These scenario descriptions are subsequently broken down into several sub-descriptions for specified details such as behaviors and locations of vehicles. The agent then distinctively transforms the textually described sub-scenarios into domain-specific languages, which then generate actual code for prediction and control in simulators, facilitating the creation of diverse and complex scenarios within the CARLA simulation environment. A key part of our agent is a comprehensive knowledge retrieval component, which efficiently translates specific textual descriptions into corresponding domain-specific code snippets by training a knowledge database containing the scenario description and code pairs. Extensive experimental results underscore the efficacy of ChatScene in improving the safety of autonomous vehicles. For instance, the scenarios generated by ChatScene show a 15% increase in collision rates compared to state-of-the-art baselines when tested against different reinforcement learning-based ego vehicles. Furthermore, we show that by using our generated safety-critical scenarios to fine-tune different RL-based autonomous driving models, they can achieve a 9% reduction in collision rates, surpassing current SOTA methods. ChatScene effectively bridges the gap between textual descriptions of traffic scenarios and practical CARLA simulations, providing a unified way to conveniently generate safety-critical scenarios for safety testing and improvement for AVs.

Jiawei Zhang, Chejian Xu, Yu Gai et al.

This paper introduces KnowHalu, a novel approach for detecting hallucinations in text generated by large language models (LLMs), utilizing step-wise reasoning, multi-formulation query, multi-form knowledge for factual checking, and fusion-based detection mechanism. As LLMs are increasingly applied across various domains, ensuring that their outputs are not hallucinated is critical. Recognizing the limitations of existing approaches that either rely on the self-consistency check of LLMs or perform post-hoc fact-checking without considering the complexity of queries or the form of knowledge, KnowHalu proposes a two-phase process for hallucination detection. In the first phase, it identifies non-fabrication hallucinations--responses that, while factually correct, are irrelevant or non-specific to the query. The second phase, multi-form based factual checking, contains five key steps: reasoning and query decomposition, knowledge retrieval, knowledge optimization, judgment generation, and judgment aggregation. Our extensive evaluations demonstrate that KnowHalu significantly outperforms SOTA baselines in detecting hallucinations across diverse tasks, e.g., improving by 15.65% in QA tasks and 5.50% in summarization tasks, highlighting its effectiveness and versatility in detecting hallucinations in LLM-generated content.

Chejian Xu, Mintong Kang, Jiawei Zhang et al.

Foundation model-based agents are increasingly used to automate complex tasks, enhancing efficiency and productivity. However, their access to sensitive resources and autonomous decision-making also introduce significant security risks, where successful attacks could lead to severe consequences. To systematically uncover these vulnerabilities, we propose AdvAgent, a black-box red-teaming framework for attacking web agents. Unlike existing approaches, AdvAgent employs a reinforcement learning-based pipeline to train an adversarial prompter model that optimizes adversarial prompts using feedback from the black-box agent. With careful attack design, these prompts effectively exploit agent weaknesses while maintaining stealthiness and controllability. Extensive evaluations demonstrate that AdvAgent achieves high success rates against state-of-the-art GPT-4-based web agents across diverse web tasks. Furthermore, we find that existing prompt-based defenses provide only limited protection, leaving agents vulnerable to our framework. These findings highlight critical vulnerabilities in current web agents and emphasize the urgent need for stronger defense mechanisms. We release code at https://ai-secure.github.io/AdvAgent/.

Mintong Kang, Chejian Xu, Bo Li

Recent advancements in large audio-language models (LALMs) have enabled speech-based user interactions, significantly enhancing user experience and accelerating the deployment of LALMs in real-world applications. However, ensuring the safety of LALMs is crucial to prevent risky outputs that may raise societal concerns or violate AI regulations. Despite the importance of this issue, research on jailbreaking LALMs remains limited due to their recent emergence and the additional technical challenges they present compared to attacks on DNN-based audio models. Specifically, the audio encoders in LALMs, which involve discretization operations, often lead to gradient shattering, hindering the effectiveness of attacks relying on gradient-based optimizations. The behavioral variability of LALMs further complicates the identification of effective (adversarial) optimization targets. Moreover, enforcing stealthiness constraints on adversarial audio waveforms introduces a reduced, non-convex feasible solution space, further intensifying the challenges of the optimization process. To overcome these challenges, we develop AdvWave, the first jailbreak framework against LALMs. We propose a dual-phase optimization method that addresses gradient shattering, enabling effective end-to-end gradient-based optimization. Additionally, we develop an adaptive adversarial target search algorithm that dynamically adjusts the adversarial optimization target based on the response patterns of LALMs for specific queries. To ensure that adversarial audio remains perceptually natural to human listeners, we design a classifier-guided optimization approach that generates adversarial noise resembling common urban sounds. Extensive evaluations on multiple advanced LALMs demonstrate that AdvWave outperforms baseline methods, achieving a 40% higher average jailbreak attack success rate.

Lingzhi Yuan, Xinfeng Li, Chejian Xu et al.

Recent text-to-image (T2I) models have exhibited remarkable performance in generating high-quality images from text descriptions. However, these models are vulnerable to misuse, particularly generating not-safe-for-work (NSFW) content, such as sexually explicit, violent, political, and disturbing images, raising serious ethical concerns. In this work, we present PromptGuard, a novel content moderation technique that draws inspiration from the system prompt mechanism in large language models (LLMs) for safety alignment. Unlike LLMs, T2I models lack a direct interface for enforcing behavioral guidelines. Our key idea is to optimize a safety soft prompt that functions as an implicit system prompt within the T2I model's textual embedding space. This universal soft prompt (P*) directly moderates NSFW inputs, enabling safe yet realistic image generation without altering the inference efficiency or requiring proxy models. We further enhance its reliability and helpfulness through a divide-and-conquer strategy, which optimizes category-specific soft prompts and combines them into holistic safety guidance. Extensive experiments across five datasets demonstrate that PromptGuard effectively mitigates NSFW content generation while preserving high-quality benign outputs. PromptGuard achieves 3.8 times faster than prior content moderation methods, surpassing eight state-of-the-art defenses with an optimal unsafe ratio down to 5.84%.

Chejian Xu, Jiawei Zhang, Zhaorun Chen et al. · berkeley

Multimodal foundation models (MMFMs) play a crucial role in various applications, including autonomous driving, healthcare, and virtual assistants. However, several studies have revealed vulnerabilities in these models, such as generating unsafe content by text-to-image models. Existing benchmarks on multimodal models either predominantly assess the helpfulness of these models, or only focus on limited perspectives such as fairness and privacy. In this paper, we present the first unified platform, MMDT (Multimodal DecodingTrust), designed to provide a comprehensive safety and trustworthiness evaluation for MMFMs. Our platform assesses models from multiple perspectives, including safety, hallucination, fairness/bias, privacy, adversarial robustness, and out-of-distribution (OOD) generalization. We have designed various evaluation scenarios and red teaming algorithms under different tasks for each perspective to generate challenging data, forming a high-quality benchmark. We evaluate a range of multimodal models using MMDT, and our findings reveal a series of vulnerabilities and areas for improvement across these perspectives. This work introduces the first comprehensive and unique safety and trustworthiness evaluation platform for MMFMs, paving the way for developing safer and more reliable MMFMs and systems. Our platform and benchmark are available at https://mmdecodingtrust.github.io/.

Chejian Xu, Wei Ping, Peng Xu et al.

Long-context capabilities are essential for a wide range of applications, including document and video understanding, in-context learning, and inference-time scaling, all of which require models to process and reason over long sequences of text and multimodal data. In this work, we introduce a efficient training recipe for building ultra-long context LLMs from aligned instruct model, pushing the boundaries of context lengths from 128K to 1M, 2M, and 4M tokens. Our approach leverages efficient continued pretraining strategies to extend the context window and employs effective instruction tuning to maintain the instruction-following and reasoning abilities. Our UltraLong-8B, built on Llama3.1-Instruct with our recipe, achieves state-of-the-art performance across a diverse set of long-context benchmarks. Importantly, models trained with our approach maintain competitive performance on standard benchmarks, demonstrating balanced improvements for both long and short context tasks. We further provide an in-depth analysis of key design choices, highlighting the impacts of scaling strategies and data composition. Our findings establish a robust framework for efficiently scaling context lengths while preserving general model capabilities. We release all model weights at: https://ultralong.github.io/.

Zijian Huang, Wenda Chu, Linyi Li et al.

Multi-sensor fusion systems (MSFs) play a vital role as the perception module in modern autonomous vehicles (AVs). Therefore, ensuring their robustness against common and realistic adversarial semantic transformations, such as rotation and shifting in the physical world, is crucial for the safety of AVs. While empirical evidence suggests that MSFs exhibit improved robustness compared to single-modal models, they are still vulnerable to adversarial semantic transformations. Despite the proposal of empirical defenses, several works show that these defenses can be attacked again by new adaptive attacks. So far, there is no certified defense proposed for MSFs. In this work, we propose the first robustness certification framework COMMIT certify robustness of multi-sensor fusion systems against semantic attacks. In particular, we propose a practical anisotropic noise mechanism that leverages randomized smoothing with multi-modal data and performs a grid-based splitting method to characterize complex semantic transformations. We also propose efficient algorithms to compute the certification in terms of object detection accuracy and IoU for large-scale MSF models. Empirically, we evaluate the efficacy of COMMIT in different settings and provide a comprehensive benchmark of certified robustness for different MSF models using the CARLA simulation platform. We show that the certification for MSF models is at most 48.39% higher than that of single-modal models, which validates the advantages of MSF models. We believe our certification framework and benchmark will contribute an important step towards certifiably robust AVs in practice.

Wenhao Ding, Chejian Xu, Mansur Arief et al.

Autonomous driving systems have witnessed a significant development during the past years thanks to the advance in machine learning-enabled sensing and decision-making algorithms. One critical challenge for their massive deployment in the real world is their safety evaluation. Most existing driving systems are still trained and evaluated on naturalistic scenarios collected from daily life or heuristically-generated adversarial ones. However, the large population of cars, in general, leads to an extremely low collision rate, indicating that the safety-critical scenarios are rare in the collected real-world data. Thus, methods to artificially generate scenarios become crucial to measure the risk and reduce the cost. In this survey, we focus on the algorithms of safety-critical scenario generation in autonomous driving. We first provide a comprehensive taxonomy of existing algorithms by dividing them into three categories: data-driven generation, adversarial generation, and knowledge-based generation. Then, we discuss useful tools for scenario generation, including simulation platforms and packages. Finally, we extend our discussion to five main challenges of current works -- fidelity, efficiency, diversity, transferability, controllability -- and research opportunities lighted up by these challenges.

Boxin Wang, Chejian Xu, Shuohang Wang et al.

Large-scale pre-trained language models have achieved tremendous success across a wide range of natural language understanding (NLU) tasks, even surpassing human performance. However, recent studies reveal that the robustness of these models can be challenged by carefully crafted textual adversarial examples. While several individual datasets have been proposed to evaluate model robustness, a principled and comprehensive benchmark is still missing. In this paper, we present Adversarial GLUE (AdvGLUE), a new multi-task benchmark to quantitatively and thoroughly explore and evaluate the vulnerabilities of modern large-scale language models under various types of adversarial attacks. In particular, we systematically apply 14 textual adversarial attack methods to GLUE tasks to construct AdvGLUE, which is further validated by humans for reliable annotations. Our findings are summarized as follows. (i) Most existing adversarial attack algorithms are prone to generating invalid or ambiguous adversarial examples, with around 90% of them either changing the original semantic meanings or misleading human annotators as well. Therefore, we perform a careful filtering process to curate a high-quality benchmark. (ii) All the language models and robust training methods we tested perform poorly on AdvGLUE, with scores lagging far behind the benign accuracy. We hope our work will motivate the development of new adversarial attacks that are more stealthy and semantic-preserving, as well as new robust language models against sophisticated adversarial attacks. AdvGLUE is available at https://adversarialglue.github.io.