Aishan Liu

Jiakai Wang, Zixin Yin, Pengfei Hu et al.

To operate in real-world high-stakes environments, deep learning systems have to endure noises that have been continuously thwarting their robustness. Data-end defense, which improves robustness by operations on input data instead of modifying models, has attracted intensive attention due to its feasibility in practice. However, previous data-end defenses show low generalization against diverse noises and weak transferability across multiple models. Motivated by the fact that robust recognition depends on both local and global features, we propose a defensive patch generation framework to address these problems by helping models better exploit these features. For the generalization against diverse noises, we inject class-specific identifiable patterns into a confined local patch prior, so that defensive patches could preserve more recognizable features towards specific classes, leading models for better recognition under noises. For the transferability across multiple models, we guide the defensive patches to capture more global feature correlations within a class, so that they could activate model-shared global perceptions and transfer better among models. Our defensive patches show great potentials to improve application robustness in practice by simply sticking them around target objects. Extensive experiments show that we outperform others by large margins (improve 20+\% accuracy for both adversarial and corruption robustness on average in the digital and physical world). Our codes are available at https://github.com/nlsde-safety-team/DefensivePatch

Simin Li, Jun Guo, Jingqiao Xiu et al.

This study probes the vulnerabilities of cooperative multi-agent reinforcement learning (c-MARL) under adversarial attacks, a critical determinant of c-MARL's worst-case performance prior to real-world implementation. Current observation-based attacks, constrained by white-box assumptions, overlook c-MARL's complex multi-agent interactions and cooperative objectives, resulting in impractical and limited attack capabilities. To address these shortcomes, we propose Adversarial Minority Influence (AMI), a practical and strong for c-MARL. AMI is a practical black-box attack and can be launched without knowing victim parameters. AMI is also strong by considering the complex multi-agent interaction and the cooperative goal of agents, enabling a single adversarial agent to unilaterally misleads majority victims to form targeted worst-case cooperation. This mirrors minority influence phenomena in social psychology. To achieve maximum deviation in victim policies under complex agent-wise interactions, our unilateral attack aims to characterize and maximize the impact of the adversary on the victims. This is achieved by adapting a unilateral agent-wise relation metric derived from mutual information, thereby mitigating the adverse effects of victim influence on the adversary. To lead the victims into a jointly detrimental scenario, our targeted attack deceives victims into a long-term, cooperatively harmful situation by guiding each victim towards a specific target, determined through a trial-and-error process executed by a reinforcement learning agent. Through AMI, we achieve the first successful attack against real-world robot swarms and effectively fool agents in simulated environments into collectively worst-case scenarios, including Starcraft II and Multi-agent Mujoco. The source code and demonstrations can be found at: https://github.com/DIG-Beihang/AMI.

Jun Guo, Aishan Liu, Xingyu Zheng et al.

Despite the broad application of Machine Learning models as a Service (MLaaS), they are vulnerable to model stealing attacks. These attacks can replicate the model functionality by using the black-box query process without any prior knowledge of the target victim model. Existing stealing defenses add deceptive perturbations to the victim's posterior probabilities to mislead the attackers. However, these defenses are now suffering problems of high inference computational overheads and unfavorable trade-offs between benign accuracy and stealing robustness, which challenges the feasibility of deployed models in practice. To address the problems, this paper proposes Isolation and Induction (InI), a novel and effective training framework for model stealing defenses. Instead of deploying auxiliary defense modules that introduce redundant inference time, InI directly trains a defensive model by isolating the adversary's training gradient from the expected gradient, which can effectively reduce the inference computational cost. In contrast to adding perturbations over model predictions that harm the benign accuracy, we train models to produce uninformative outputs against stealing queries, which can induce the adversary to extract little useful knowledge from victim models with minimal impact on the benign performance. Extensive experiments on several visual classification datasets (e.g., MNIST and CIFAR10) demonstrate the superior robustness (up to 48% reduction on stealing accuracy) and speed (up to 25.4x faster) of our InI over other state-of-the-art methods. Our codes can be found in https://github.com/DIG-Beihang/InI-Model-Stealing-Defense.

Simin Li, Huangxinxin Xu, Jiakai Wang et al.

Billions of people are sharing their daily life images on social media every day. However, their biometric information (e.g., fingerprint) could be easily stolen from these images. The threat of fingerprint leakage from social media raises a strong desire for anonymizing shared images while maintaining image qualities, since fingerprints act as a lifelong individual biometric password. To guard the fingerprint leakage, adversarial attack emerges as a solution by adding imperceptible perturbations on images. However, existing works are either weak in black-box transferability or appear unnatural. Motivated by visual perception hierarchy (i.e., high-level perception exploits model-shared semantics that transfer well across models while low-level perception extracts primitive stimulus and will cause high visual sensitivities given suspicious stimulus), we propose FingerSafe, a hierarchical perceptual protective noise injection framework to address the mentioned problems. For black-box transferability, we inject protective noises on fingerprint orientation field to perturb the model-shared high-level semantics (i.e., fingerprint ridges). Considering visual naturalness, we suppress the low-level local contrast stimulus by regularizing the response of Lateral Geniculate Nucleus. Our FingerSafe is the first to provide feasible fingerprint protection in both digital (up to 94.12%) and realistic scenarios (Twitter and Facebook, up to 68.75%). Our code can be found at https://github.com/nlsde-safety-team/FingerSafe.

Jiawei Liang, Siyuan Liang, Aishan Liu et al.

Knowledge Distillation (KD) for object detection aims to train a compact detector by transferring knowledge from a teacher model. Since the teacher model perceives data in a way different from humans, existing KD methods only distill knowledge that is consistent with labels annotated by human expert while neglecting knowledge that is not consistent with human perception, which results in insufficient distillation and sub-optimal performance. In this paper, we propose inconsistent knowledge distillation (IKD), which aims to distill knowledge inherent in the teacher model's counter-intuitive perceptions. We start by considering the teacher model's counter-intuitive perceptions of frequency and non-robust features. Unlike previous works that exploit fine-grained features or introduce additional regularizations, we extract inconsistent knowledge by providing diverse input using data augmentation. Specifically, we propose a sample-specific data augmentation to transfer the teacher model's ability in capturing distinct frequency components and suggest an adversarial feature augmentation to extract the teacher model's perceptions of non-robust features in the data. Extensive experiments demonstrate the effectiveness of our method which outperforms state-of-the-art KD baselines on one-stage, two-stage and anchor-free object detectors (at most +1.0 mAP). Our codes will be made available at \url{https://github.com/JWLiang007/IKD.git}.

Xin Dong, Rui Wang, Siyuan Liang et al.

Billions of people are sharing their daily live images on social media everyday. However, malicious collectors use deep face recognition systems to easily steal their biometric information (e.g., faces) from these images. Some studies are being conducted to generate encrypted face photos using adversarial attacks by introducing imperceptible perturbations to reduce face information leakage. However, existing studies need stronger black-box scenario feasibility and more natural visual appearances, which challenge the feasibility of privacy protection. To address these problems, we propose a frequency-restricted identity-agnostic (FRIA) framework to encrypt face images from unauthorized face recognition without access to personal information. As for the weak black-box scenario feasibility, we obverse that representations of the average feature in multiple face recognition models are similar, thus we propose to utilize the average feature via the crawled dataset from the Internet as the target to guide the generation, which is also agnostic to identities of unknown face recognition systems; in nature, the low-frequency perturbations are more visually perceptible by the human vision system. Inspired by this, we restrict the perturbation in the low-frequency facial regions by discrete cosine transform to achieve the visual naturalness guarantee. Extensive experiments on several face recognition models demonstrate that our FRIA outperforms other state-of-the-art methods in generating more natural encrypted faces while attaining high black-box attack success rates of 96%. In addition, we validate the efficacy of FRIA using real-world black-box commercial API, which reveals the potential of FRIA in practice. Our codes can be found in https://github.com/XinDong10/FRIA.

Haotong Qin, Yifu Ding, Mingyuan Zhang et al.

The large pre-trained BERT has achieved remarkable performance on Natural Language Processing (NLP) tasks but is also computation and memory expensive. As one of the powerful compression approaches, binarization extremely reduces the computation and memory consumption by utilizing 1-bit parameters and bitwise operations. Unfortunately, the full binarization of BERT (i.e., 1-bit weight, embedding, and activation) usually suffer a significant performance drop, and there is rare study addressing this problem. In this paper, with the theoretical justification and empirical analysis, we identify that the severe performance drop can be mainly attributed to the information degradation and optimization direction mismatch respectively in the forward and backward propagation, and propose BiBERT, an accurate fully binarized BERT, to eliminate the performance bottlenecks. Specifically, BiBERT introduces an efficient Bi-Attention structure for maximizing representation information statistically and a Direction-Matching Distillation (DMD) scheme to optimize the full binarized BERT accurately. Extensive experiments show that BiBERT outperforms both the straightforward baseline and existing state-of-the-art quantized BERTs with ultra-low bit activations by convincing margins on the NLP benchmark. As the first fully binarized BERT, our method yields impressive 56.3 times and 31.2 times saving on FLOPs and model size, demonstrating the vast advantages and potential of the fully binarized BERT model in real-world resource-constrained scenarios.

Yan Wang, Yuhang Li, Ruihao Gong et al.

Extensive studies have shown that deep learning models are vulnerable to adversarial and natural noises, yet little is known about model robustness on noises caused by different system implementations. In this paper, we for the first time introduce SysNoise, a frequently occurred but often overlooked noise in the deep learning training-deployment cycle. In particular, SysNoise happens when the source training system switches to a disparate target system in deployments, where various tiny system mismatch adds up to a non-negligible difference. We first identify and classify SysNoise into three categories based on the inference stage; we then build a holistic benchmark to quantitatively measure the impact of SysNoise on 20+ models, comprehending image classification, object detection, instance segmentation and natural language processing tasks. Our extensive experiments revealed that SysNoise could bring certain impacts on model robustness across different tasks and common mitigations like data augmentation and adversarial training show limited effects on it. Together, our findings open a new research topic and we hope this work will raise research attention to deep learning deployment systems accounting for model performance. We have open-sourced the benchmark and framework at https://modeltc.github.io/systemnoise_web.

Aishan Liu, Jun Guo, Jiakai Wang et al.

Adversarial attacks are valuable for evaluating the robustness of deep learning models. Existing attacks are primarily conducted on the visible light spectrum (e.g., pixel-wise texture perturbation). However, attacks targeting texture-free X-ray images remain underexplored, despite the widespread application of X-ray imaging in safety-critical scenarios such as the X-ray detection of prohibited items. In this paper, we take the first step toward the study of adversarial attacks targeted at X-ray prohibited item detection, and reveal the serious threats posed by such attacks in this safety-critical scenario. Specifically, we posit that successful physical adversarial attacks in this scenario should be specially designed to circumvent the challenges posed by color/texture fading and complex overlapping. To this end, we propose X-adv to generate physically printable metals that act as an adversarial agent capable of deceiving X-ray detectors when placed in luggage. To resolve the issues associated with color/texture fading, we develop a differentiable converter that facilitates the generation of 3D-printable objects with adversarial shapes, using the gradients of a surrogate model rather than directly generating adversarial textures. To place the printed 3D adversarial objects in luggage with complex overlapped instances, we design a policy-based reinforcement learning strategy to find locations eliciting strong attack performance in worst-case scenarios whereby the prohibited items are heavily occluded by other items. To verify the effectiveness of the proposed X-Adv, we conduct extensive experiments in both the digital and the physical world (employing a commercial X-ray security inspection system for the latter case). Furthermore, we present the physical-world X-ray adversarial attack dataset XAD.

Kunsheng Tang, Wenbo Zhou, Jie Zhang et al.

Large language models (LLMs) have exhibited remarkable capabilities in natural language generation, but they have also been observed to magnify societal biases, particularly those related to gender. In response to this issue, several benchmarks have been proposed to assess gender bias in LLMs. However, these benchmarks often lack practical flexibility or inadvertently introduce biases. To address these shortcomings, we introduce GenderCARE, a comprehensive framework that encompasses innovative Criteria, bias Assessment, Reduction techniques, and Evaluation metrics for quantifying and mitigating gender bias in LLMs. To begin, we establish pioneering criteria for gender equality benchmarks, spanning dimensions such as inclusivity, diversity, explainability, objectivity, robustness, and realisticity. Guided by these criteria, we construct GenderPair, a novel pair-based benchmark designed to assess gender bias in LLMs comprehensively. Our benchmark provides standardized and realistic evaluations, including previously overlooked gender groups such as transgender and non-binary individuals. Furthermore, we develop effective debiasing techniques that incorporate counterfactual data augmentation and specialized fine-tuning strategies to reduce gender bias in LLMs without compromising their overall performance. Extensive experiments demonstrate a significant reduction in various gender bias benchmarks, with reductions peaking at over 90% and averaging above 35% across 17 different LLMs. Importantly, these reductions come with minimal variability in mainstream language tasks, remaining below 2%. By offering a realistic assessment and tailored reduction of gender biases, we hope that our GenderCARE can represent a significant step towards achieving fairness and equity in LLMs. More details are available at https://github.com/kstanghere/GenderCARE-ccs24.

Jian Yang, Wei Zhang, Jiajun Wu et al.

Recent code large language models have achieved remarkable progress on general programming tasks. Nevertheless, their performance degrades significantly in industrial scenarios that require reasoning about hardware semantics, specialized language constructs, and strict resource constraints. To address these challenges, we introduce InCoder-32B (Industrial-Coder-32B), the first 32B-parameter code foundation model unifying code intelligence across chip design, GPU kernel optimization, embedded systems, compiler optimization, and 3D modeling. By adopting an efficient architecture, we train InCoder-32B from scratch with general code pre-training, curated industrial code annealing, mid-training that progressively extends context from 8K to 128K tokens with synthetic industrial reasoning data, and post-training with execution-grounded verification. We conduct extensive evaluation on 14 mainstream general code benchmarks and 9 industrial benchmarks spanning 4 specialized domains. Results show InCoder-32B achieves highly competitive performance on general tasks while establishing strong open-source baselines across industrial domains.

Qitong Sun, Jun Han, Tianlin Li et al.

Improving GPU kernel efficiency is crucial for advancing AI systems. Recent work has explored leveraging large language models (LLMs) for GPU kernel generation and optimization. However, existing LLM-based kernel optimization pipelines typically rely on opaque, implicitly learned heuristics within the LLMs to determine optimization strategies. This leads to inefficient trial-and-error and weakly interpretable optimizations. Our key insight is to replace implicit heuristics with expert optimization skills that are knowledge-driven and aware of task trajectories. Specifically, we present KernelSkill, a multi-agent framework with a dual-level memory architecture. KernelSkill operates by coordinating agents with long-term memory of reusable expert skills and short-term memory to prevent repetitive backtracking. On KernelBench Levels 1-3, KernelSkill achieves a 100% success rate and average speedups of 5.44x, 2.82x, and 1.92x over Torch Eager on Levels 1, 2, and 3, respectively, outperforming prior baselines. Code is available at https://github.com/0satan0/KernelMem/.

Yisong Xiao, Aishan Liu, Tianyuan Zhang et al.

Quantization has emerged as an essential technique for deploying deep neural networks (DNNs) on devices with limited resources. However, quantized models exhibit vulnerabilities when exposed to various noises in real-world applications. Despite the importance of evaluating the impact of quantization on robustness, existing research on this topic is limited and often disregards established principles of robustness evaluation, resulting in incomplete and inconclusive findings. To address this gap, we thoroughly evaluated the robustness of quantized models against various noises (adversarial attacks, natural corruptions, and systematic noises) on ImageNet. The comprehensive evaluation results empirically provide valuable insights into the robustness of quantized models in various scenarios, for example: (1) quantized models exhibit higher adversarial robustness than their floating-point counterparts, but are more vulnerable to natural corruptions and systematic noises; (2) in general, increasing the quantization bit-width results in a decrease in adversarial robustness, an increase in natural robustness, and an increase in systematic robustness; (3) among corruption methods, \textit{impulse noise} and \textit{glass blur} are the most harmful to quantized models, while \textit{brightness} has the least impact; (4) among systematic noises, the \textit{nearest neighbor interpolation} has the highest impact, while bilinear interpolation, cubic interpolation, and area interpolation are the three least harmful. Our research contributes to advancing the robust quantization of models and their deployment in real-world scenarios.

Aishan Liu, Shiyu Tang, Siyuan Liang et al.

Adversarial training has been demonstrated to be one of the most effective remedies for defending adversarial examples, yet it often suffers from the huge robustness generalization gap on unseen testing adversaries, deemed as the adversarially robust generalization problem. Despite the preliminary understandings devoted to adversarially robust generalization, little is known from the architectural perspective. To bridge the gap, this paper for the first time systematically investigated the relationship between adversarially robust generalization and architectural design. Inparticular, we comprehensively evaluated 20 most representative adversarially trained architectures on ImageNette and CIFAR-10 datasets towards multiple `p-norm adversarial attacks. Based on the extensive experiments, we found that, under aligned settings, Vision Transformers (e.g., PVT, CoAtNet) often yield better adversarially robust generalization while CNNs tend to overfit on specific attacks and fail to generalize on multiple adversaries. To better understand the nature behind it, we conduct theoretical analysis via the lens of Rademacher complexity. We revealed the fact that the higher weight sparsity contributes significantly towards the better adversarially robust generalization of Transformers, which can be often achieved by the specially-designed attention blocks. We hope our paper could help to better understand the mechanism for designing robust DNNs. Our model weights can be found at http://robust.art.

Chunyu Sun, Chenye Xu, Chengyuan Yao et al.

Adversarial training (AT) methods are effective against adversarial attacks, yet they introduce severe disparity of accuracy and robustness between different classes, known as the robust fairness problem. Previously proposed Fair Robust Learning (FRL) adaptively reweights different classes to improve fairness. However, the performance of the better-performed classes decreases, leading to a strong performance drop. In this paper, we observed two unfair phenomena during adversarial training: different difficulties in generating adversarial examples from each class (source-class fairness) and disparate target class tendencies when generating adversarial examples (target-class fairness). From the observations, we propose Balance Adversarial Training (BAT) to address the robust fairness problem. Regarding source-class fairness, we adjust the attack strength and difficulties of each class to generate samples near the decision boundary for easier and fairer model learning; considering target-class fairness, by introducing a uniform distribution constraint, we encourage the adversarial example generation process for each class with a fair tendency. Extensive experiments conducted on multiple datasets (CIFAR-10, CIFAR-100, and ImageNette) demonstrate that our method can significantly outperform other baselines in mitigating the robust fairness problem (+5-10\% on the worst class accuracy)

Siyuan Liang, Mingli Zhu, Aishan Liu et al.

Studying backdoor attacks is valuable for model copyright protection and enhancing defenses. While existing backdoor attacks have successfully infected multimodal contrastive learning models such as CLIP, they can be easily countered by specialized backdoor defenses for MCL models. This paper reveals the threats in this practical scenario that backdoor attacks can remain effective even after defenses and introduces the \emph{\toolns} attack, which is resistant to backdoor detection and model fine-tuning defenses. To achieve this, we draw motivations from the perspective of the Bayesian rule and propose a dual-embedding guided framework for backdoor attacks. Specifically, we ensure that visual trigger patterns approximate the textual target semantics in the embedding space, making it challenging to detect the subtle parameter variations induced by backdoor learning on such natural trigger patterns. Additionally, we optimize the visual trigger patterns to align the poisoned samples with target vision features in order to hinder the backdoor unlearning through clean fine-tuning. Extensive experiments demonstrate that our attack significantly outperforms state-of-the-art baselines (+45.3% ASR) in the presence of SoTA backdoor defenses, rendering these mitigation and detection strategies virtually ineffective. Furthermore, our approach effectively attacks some more rigorous scenarios like downstream tasks. We believe that this paper raises awareness regarding the potential threats associated with the practical application of multimodal contrastive learning and encourages the development of more robust defense mechanisms.

Aishan Liu, Yuguang Zhou, Xianglong Liu et al.

Large language models (LLMs) have transformed the development of embodied intelligence. By providing a few contextual demonstrations, developers can utilize the extensive internal knowledge of LLMs to effortlessly translate complex tasks described in abstract language into sequences of code snippets, which will serve as the execution logic for embodied agents. However, this paper uncovers a significant backdoor security threat within this process and introduces a novel method called \method{}. By poisoning just a few contextual demonstrations, attackers can covertly compromise the contextual environment of a black-box LLM, prompting it to generate programs with context-dependent defects. These programs appear logically sound but contain defects that can activate and induce unintended behaviors when the operational agent encounters specific triggers in its interactive environment. To compromise the LLM's contextual environment, we employ adversarial in-context generation to optimize poisoned demonstrations, where an LLM judge evaluates these poisoned prompts, reporting to an additional LLM that iteratively optimizes the demonstration in a two-player adversarial game using chain-of-thought reasoning. To enable context-dependent behaviors in downstream agents, we implement a dual-modality activation strategy that controls both the generation and execution of program defects through textual and visual triggers. We expand the scope of our attack by developing five program defect modes that compromise key aspects of confidentiality, integrity, and availability in embodied agents. To validate the effectiveness of our approach, we conducted extensive experiments across various tasks, including robot planning, robot manipulation, and compositional visual reasoning. Additionally, we demonstrate the potential impact of our approach by successfully attacking real-world autonomous driving systems.

Hangtao Zhang, Chenyu Zhu, Xianlong Wang et al.

Embodied AI represents systems where AI is integrated into physical entities. Large Language Model (LLM), which exhibits powerful language understanding abilities, has been extensively employed in embodied AI by facilitating sophisticated task planning. However, a critical safety issue remains overlooked: could these embodied LLMs perpetrate harmful behaviors? In response, we introduce BadRobot, a novel attack paradigm aiming to make embodied LLMs violate safety and ethical constraints through typical voice-based user-system interactions. Specifically, three vulnerabilities are exploited to achieve this type of attack: (i) manipulation of LLMs within robotic systems, (ii) misalignment between linguistic outputs and physical actions, and (iii) unintentional hazardous behaviors caused by world knowledge's flaws. Furthermore, we construct a benchmark of various malicious physical action queries to evaluate BadRobot's attack performance. Based on this benchmark, extensive experiments against existing prominent embodied LLM frameworks (e.g., Voxposer, Code as Policies, and ProgPrompt) demonstrate the effectiveness of our BadRobot.

Tianlin Li, Qing Guo, Aishan Liu et al.

Deep neural networks (DNNs) have made significant progress, but often suffer from fairness issues, as deep models typically show distinct accuracy differences among certain subgroups (e.g., males and females). Existing research addresses this critical issue by employing fairness-aware loss functions to constrain the last-layer outputs and directly regularize DNNs. Although the fairness of DNNs is improved, it is unclear how the trained network makes a fair prediction, which limits future fairness improvements. In this paper, we investigate fairness from the perspective of decision rationale and define the parameter parity score to characterize the fair decision process of networks by analyzing neuron influence in various subgroups. Extensive empirical studies show that the unfair issue could arise from the unaligned decision rationales of subgroups. Existing fairness regularization terms fail to achieve decision rationale alignment because they only constrain last-layer outputs while ignoring intermediate neuron alignment. To address the issue, we formulate the fairness as a new task, i.e., decision rationale alignment that requires DNNs' neurons to have consistent responses on subgroups at both intermediate processes and the final prediction. To make this idea practical during optimization, we relax the naive objective function and propose gradient-guided parity alignment, which encourages gradient-weighted consistency of neurons across subgroups. Extensive experiments on a variety of datasets show that our method can significantly enhance fairness while sustaining a high level of accuracy and outperforming other approaches by a wide margin.

Yuhang Wang, Huafeng Shi, Rui Min et al.

Extensive evidence has demonstrated that deep neural networks (DNNs) are vulnerable to backdoor attacks, which motivates the development of backdoor attacks detection. Most detection methods are designed to verify whether a model is infected with presumed types of backdoor attacks, yet the adversary is likely to generate diverse backdoor attacks in practice that are unforeseen to defenders, which challenge current detection strategies. In this paper, we focus on this more challenging scenario and propose a universal backdoor attacks detection method named Adaptive Adversarial Probe (A2P). Specifically, we posit that the challenge of universal backdoor attacks detection lies in the fact that different backdoor attacks often exhibit diverse characteristics in trigger patterns (i.e., sizes and transparencies). Therefore, our A2P adopts a global-to-local probing framework, which adversarially probes images with adaptive regions/budgets to fit various backdoor triggers of different sizes/transparencies. Regarding the probing region, we propose the attention-guided region generation strategy that generates region proposals with different sizes/locations based on the attention of the target model, since trigger regions often manifest higher model activation. Considering the attack budget, we introduce the box-to-sparsity scheduling that iteratively increases the perturbation budget from box to sparse constraint, so that we could better activate different latent backdoors with different transparencies. Extensive experiments on multiple datasets (CIFAR-10, GTSRB, Tiny-ImageNet) demonstrate that our method outperforms state-of-the-art baselines by large margins (+12%).

Jiakai Wang, Xianglong Liu, Jin Hu et al.

Deep neural networks (DNNs) have demonstrated high vulnerability to adversarial examples, raising broad security concerns about their applications. Besides the attacks in the digital world, the practical implications of adversarial examples in the physical world present significant challenges and safety concerns. However, current research on physical adversarial examples (PAEs) lacks a comprehensive understanding of their unique characteristics, leading to limited significance and understanding. In this paper, we address this gap by thoroughly examining the characteristics of PAEs within a practical workflow encompassing training, manufacturing, and re-sampling processes. By analyzing the links between physical adversarial attacks, we identify manufacturing and re-sampling as the primary sources of distinct attributes and particularities in PAEs. Leveraging this knowledge, we develop a comprehensive analysis and classification framework for PAEs based on their specific characteristics, covering over 100 studies on physical-world adversarial examples. Furthermore, we investigate defense strategies against PAEs and identify open challenges and opportunities for future research. We aim to provide a fresh, thorough, and systematic understanding of PAEs, thereby promoting the development of robust adversarial learning and its application in open-world scenarios to provide the community with a continuously updated list of physical world adversarial sample resources, including papers, code, \etc, within the proposed framework

Xianlong Wang, Hewen Pan, Hangtao Zhang et al.

Robotic manipulation policies are increasingly empowered by \textit{large language models} (LLMs) and \textit{vision-language models} (VLMs), leveraging their understanding and perception capabilities. Recently, inference-time attacks against robotic manipulation have been extensively studied, yet backdoor attacks targeting model supply chain security in robotic policies remain largely unexplored. To fill this gap, we propose \texttt{TrojanRobot}, a backdoor injection framework for model supply chain attack scenarios, which embeds a malicious module into modular robotic policies via backdoor relationships to manipulate the LLM-to-VLM pathway and compromise the system. Our vanilla design instantiates this module as a backdoor-finetuned VLM. To further enhance attack performance, we propose a prime scheme by introducing the concept of \textit{LVLM-as-a-backdoor}, which leverages \textit{in-context instruction learning} (ICIL) to steer \textit{large vision-language model} (LVLM) behavior through backdoored system prompts. Moreover, we develop three types of prime attacks, \textit{permutation}, \textit{stagnation}, and \textit{intentional}, achieving flexible backdoor attack effects. Extensive physical-world and simulator experiments on 18 real-world manipulation tasks and 4 VLMs verify the superiority of proposed \texttt{TrojanRobot}

Tianyuan Zhang, Lu Wang, Jiaqi Kang et al.

Recent advances in deep learning have markedly improved autonomous driving (AD) models, particularly end-to-end systems that integrate perception, prediction, and planning stages, achieving state-of-the-art performance. However, these models remain vulnerable to adversarial attacks, where human-imperceptible perturbations can disrupt decision-making processes. While adversarial training is an effective method for enhancing model robustness against such attacks, no prior studies have focused on its application to end-to-end AD models. In this paper, we take the first step in adversarial training for end-to-end AD models and present a novel Module-wise Adaptive Adversarial Training (MA2T). However, extending conventional adversarial training to this context is highly non-trivial, as different stages within the model have distinct objectives and are strongly interconnected. To address these challenges, MA2T first introduces Module-wise Noise Injection, which injects noise before the input of different modules, targeting training models with the guidance of overall objectives rather than each independent module loss. Additionally, we introduce Dynamic Weight Accumulation Adaptation, which incorporates accumulated weight changes to adaptively learn and adjust the loss weights of each module based on their contributions (accumulated reduction rates) for better balance and robust training. To demonstrate the efficacy of our defense, we conduct extensive experiments on the widely-used nuScenes dataset across several end-to-end AD models under both white-box and black-box attacks, where our method outperforms other baselines by large margins (+5-10%). Moreover, we validate the robustness of our defense through closed-loop evaluation in the CARLA simulation environment, showing improved resilience even against natural corruption.

Jian Yang, Shawn Guo, Lin Jing et al.

Code large language models (Code LLMs) are powerful but costly to train, with scaling laws predicting performance from model size, data, and compute. However, different programming languages (PLs) have varying impacts during pre-training that significantly affect base model performance, leading to inaccurate performance prediction. Besides, existing works focus on language-agnostic settings, neglecting the inherently multilingual nature of modern software development. Therefore, it is first necessary to investigate the scaling laws of different PLs, and then consider their mutual influences to arrive at the final multilingual scaling law. In this paper, we present the first systematic exploration of scaling laws for multilingual code pre-training, conducting over 1000+ experiments (Equivalent to 336,000+ H800 hours) across multiple PLs, model sizes (0.2B to 14B parameters), and dataset sizes (1T tokens). We establish comprehensive scaling laws for code LLMs across multiple PLs, revealing that interpreted languages (e.g., Python) benefit more from increased model size and data than compiled languages (e.g., Rust). The study demonstrates that multilingual pre-training provides synergistic benefits, particularly between syntactically similar PLs. Further, the pre-training strategy of the parallel pairing (concatenating code snippets with their translations) significantly enhances cross-lingual abilities with favorable scaling properties. Finally, a proportion-dependent multilingual scaling law is proposed to optimally allocate training tokens by prioritizing high-utility PLs (e.g., Python), balancing high-synergy pairs (e.g., JavaScript-TypeScript), and reducing allocation to fast-saturating languages (Rust), achieving superior average performance across all PLs compared to uniform distribution under the same compute budget.

Jian Yang, Wei Zhang, Yizhi Li et al.

Large language models (LLMs) have made significant strides in code generation, achieving impressive capabilities in synthesizing code snippets from natural language instructions. However, a critical challenge remains in ensuring LLMs generate factually accurate responses about programming concepts, technical implementations, etc. Most previous code-related benchmarks focus on code execution correctness, overlooking the factual accuracy of programming knowledge. To address this gap, we present CodeSimpleQA, a comprehensive bilingual benchmark designed to evaluate the factual accuracy of code LLMs in answering code-related questions, which contains carefully curated question-answer pairs in both English and Chinese, covering diverse programming languages and major computer science domains. Further, we create CodeSimpleQA-Instruct, a large-scale instruction corpus with 66M samples, and develop a post-training framework combining supervised fine-tuning and reinforcement learning. Our comprehensive evaluation of diverse LLMs reveals that even frontier LLMs struggle with code factuality. Our proposed framework demonstrates substantial improvements over the base model, underscoring the critical importance of factuality-aware alignment in developing reliable code LLMs.

Zonghao Ying, Deyue Zhang, Zonglei Jing et al.

Multi-turn jailbreak attacks simulate real-world human interactions by engaging large language models (LLMs) in iterative dialogues, exposing critical safety vulnerabilities. However, existing methods often struggle to balance semantic coherence with attack effectiveness, resulting in either benign semantic drift or ineffective detection evasion. To address this challenge, we propose Reasoning-Augmented Conversation, a novel multi-turn jailbreak framework that reformulates harmful queries into benign reasoning tasks and leverages LLMs' strong reasoning capabilities to compromise safety alignment. Specifically, we introduce an attack state machine framework to systematically model problem translation and iterative reasoning, ensuring coherent query generation across multiple turns. Building on this framework, we design gain-guided exploration, self-play, and rejection feedback modules to preserve attack semantics, enhance effectiveness, and sustain reasoning-driven attack progression. Extensive experiments on multiple LLMs demonstrate that RACE achieves state-of-the-art attack effectiveness in complex conversational scenarios, with attack success rates (ASRs) increasing by up to 96%. Notably, our approach achieves ASRs of 82% and 92% against leading commercial models, OpenAI o1 and DeepSeek R1, underscoring its potency. We release our code at https://github.com/NY1024/RACE to facilitate further research in this critical domain.

Jiawei Liang, Siyuan Liang, Aishan Liu et al.

The proliferation of face forgery techniques has raised significant concerns within society, thereby motivating the development of face forgery detection methods. These methods aim to distinguish forged faces from genuine ones and have proven effective in practical applications. However, this paper introduces a novel and previously unrecognized threat in face forgery detection scenarios caused by backdoor attack. By embedding backdoors into models and incorporating specific trigger patterns into the input, attackers can deceive detectors into producing erroneous predictions for forged faces. To achieve this goal, this paper proposes \emph{Poisoned Forgery Face} framework, which enables clean-label backdoor attacks on face forgery detectors. Our approach involves constructing a scalable trigger generator and utilizing a novel convolving process to generate translation-sensitive trigger patterns. Moreover, we employ a relative embedding method based on landmark-based regions to enhance the stealthiness of the poisoned samples. Consequently, detectors trained on our poisoned samples are embedded with backdoors. Notably, our approach surpasses SoTA backdoor baselines with a significant improvement in attack success rate (+16.39\% BD-AUC) and reduction in visibility (-12.65\% $L_\infty$). Furthermore, our attack exhibits promising performance against backdoor defenses. We anticipate that this paper will draw greater attention to the potential threats posed by backdoor attacks in face forgery detection scenarios. Our codes will be made available at \url{https://github.com/JWLiang007/PFF}

Zonghao Ying, Guangyi Zheng, Yongxin Huang et al.

This study presents the first comprehensive safety evaluation of the DeepSeek models, focusing on evaluating the safety risks associated with their generated content. Our evaluation encompasses DeepSeek's latest generation of large language models, multimodal large language models, and text-to-image models, systematically examining their performance regarding unsafe content generation. Notably, we developed a bilingual (Chinese-English) safety evaluation dataset tailored to Chinese sociocultural contexts, enabling a more thorough evaluation of the safety capabilities of Chinese-developed models. Experimental results indicate that despite their strong general capabilities, DeepSeek models exhibit significant safety vulnerabilities across multiple risk dimensions, including algorithmic discrimination and sexual content. These findings provide crucial insights for understanding and improving the safety of large foundation models. Our code is available at https://github.com/NY1024/DeepSeek-Safety-Eval.

Fanglin Xu, Wei Zhang, Jian Yang et al.

The rapid advancement of code large language models (LLMs) has sparked significant research interest in systematically evaluating their code generation capabilities, yet existing benchmarks predominantly assess models at a single structural granularity and focus on limited programming languages, obscuring fine-grained capability variations across different code scopes and multilingual scenarios. We introduce M2G-Eval, a multi-granularity, multilingual framework for evaluating code generation in large language models (LLMs) across four levels: Class, Function, Block, and Line. Spanning 18 programming languages, M2G-Eval includes 17K+ training tasks and 1,286 human-annotated, contamination-controlled test instances. We develop M2G-Eval-Coder models by training Qwen3-8B with supervised fine-tuning and Group Relative Policy Optimization. Evaluating 30 models (28 state-of-the-art LLMs plus our two M2G-Eval-Coder variants) reveals three main findings: (1) an apparent difficulty hierarchy, with Line-level tasks easiest and Class-level most challenging; (2) widening performance gaps between full- and partial-granularity languages as task complexity increases; and (3) strong cross-language correlations, suggesting that models learn transferable programming concepts. M2G-Eval enables fine-grained diagnosis of code generation capabilities and highlights persistent challenges in synthesizing complex, long-form code.

Siyuan Liang, Wei Wang, Ruoyu Chen et al.

With the emergence of foundation models, deep learning-based object detectors have shown practical usability in closed set scenarios. However, for real-world tasks, object detectors often operate in open environments, where crucial factors (e.g., data distribution, objective) that influence model learning are often changing. The dynamic and intricate nature of the open environment poses novel and formidable challenges to object detectors. Unfortunately, current research on object detectors in open environments lacks a comprehensive analysis of their distinctive characteristics, challenges, and corresponding solutions, which hinders their secure deployment in critical real-world scenarios. This paper aims to bridge this gap by conducting a comprehensive review and analysis of object detectors in open environments. We initially identified limitations of key structural components within the existing detection pipeline and propose the open environment object detector challenge framework that includes four quadrants (i.e., out-of-domain, out-of-category, robust learning, and incremental learning) based on the dimensions of the data / target changes. For each quadrant of challenges in the proposed framework, we present a detailed description and systematic analysis of the overarching goals and core difficulties, systematically review the corresponding solutions, and benchmark their performance over multiple widely adopted datasets. In addition, we engage in a discussion of open problems and potential avenues for future research. This paper aims to provide a fresh, comprehensive, and systematic understanding of the challenges and solutions associated with open-environment object detectors, thus catalyzing the development of more solid applications in real-world scenarios. A project related to this survey can be found at https://github.com/LiangSiyuan21/OEOD_Survey.

Zonghao Ying, Xiao Yang, Siyang Wu et al.

The rapid evolution of Large Language Models (LLMs) into autonomous, tool-calling agents has fundamentally altered the cybersecurity landscape. Frameworks like OpenClaw grant AI systems operating-system-level permissions and the autonomy to execute complex workflows. This level of access creates unprecedented security challenges. Consequently, traditional content-filtering defenses have become obsolete. This report presents a comprehensive security analysis of the OpenClaw ecosystem. We systematically investigate its current threat landscape, highlighting critical vulnerabilities such as prompt injection-driven Remote Code Execution (RCE), sequential tool attack chains, context amnesia, and supply chain contamination. To systematically contextualize these threats, we propose a novel tri-layered risk taxonomy for autonomous Agents, categorizing vulnerabilities across AI Cognitive, Software Execution, and Information System dimensions. To address these systemic architectural flaws, we introduce the Full-Lifecycle Agent Security Architecture (FASA). This theoretical defense blueprint advocates for zero-trust agentic execution, dynamic intent verification, and cross-layer reasoning-action correlation. Building on this framework, we present Project ClawGuard, our ongoing engineering initiative. This project aims to implement the FASA paradigm and transition autonomous agents from high-risk experimental utilities into trustworthy systems. Our code and dataset are available at https://github.com/NY1024/ClawGuard.

Yunxiang Wei, Tianlin Li, Yuwei Zheng et al.

Large language models (LLMs) have demonstrated remarkable capabilities in generating programs from natural language descriptions, yet ensuring their correctness without an external oracle remains a critical challenge. To solve the challenge, existing methods often rely on uncertainty estimation, measuring the consistency of semantics or execution behaviors across multiple samples generated by a single model. However, we observe that a single model can often converge to a consistent but incorrect solution, rendering such consistency-based proxies ineffective. To address this, we propose Ensemble Semantic Entropy (ESE), which estimates uncertainty by evaluating the consistency of samples aggregated across an ensemble of models. Experiments on LiveCodeBench demonstrate that ESE correlates more strongly with program correctness than single-model semantic entropy. Notably, in selective generation tasks with strict false-positive rate constraints, ESE improves prediction accuracy by 53.4%. Furthermore, by leveraging ESE as the decision signal, we propose a cascading test-time scaling framework Cas, which maintains performance while reducing FLOPs by 64.9% compared to single-model scaling, offering a new perspective on balancing parameter and inference scaling.

Lu Wang, Tianyuan Zhang, Yang Qu et al.

Vision-language models (VLMs) have significantly advanced autonomous driving (AD) by enhancing reasoning capabilities; however, these models remain highly susceptible to adversarial attacks. While existing research has explored white-box attacks to some extent, the more practical and challenging black-box scenarios remain largely underexplored due to their inherent difficulty. In this paper, we take the first step toward designing black-box adversarial attacks specifically targeting VLMs in AD. We identify two key challenges for achieving effective black-box attacks in this context: the effectiveness across driving reasoning chains in AD systems and the dynamic nature of driving scenarios. To address this, we propose Cascading Adversarial Disruption (CAD). It first introduces Decision Chain Disruption, which targets low-level reasoning breakdown by generating and injecting deceptive semantics, ensuring the perturbations remain effective across the entire decision-making chain. Building on this, we present Risky Scene Induction, which addresses dynamic adaptation by leveraging a surrogate VLM to understand and construct high-level risky scenarios that are likely to result in critical errors in the current driving contexts. Extensive experiments conducted on multiple AD VLMs and benchmarks demonstrate that CAD achieves state-of-the-art attack effectiveness, significantly outperforming existing methods (+13.43% on average). Moreover, we validate its practical applicability through real-world attacks on AD vehicles powered by VLMs, where the route completion rate drops by 61.11% and the vehicle crashes directly into the obstacle vehicle with adversarial patches. Finally, we release CADA dataset, comprising 18,808 adversarial visual-question-answer pairs, to facilitate further evaluation and research in this critical domain. Our codes and dataset will be available after paper's acceptance.

Zonglei Jing, Zonghao Ying, Le Wang et al.

The development of text-to-image (T2I) generative models, that enable the creation of high-quality synthetic images from textual prompts, has opened new frontiers in creative design and content generation. However, this paper reveals a significant and previously unrecognized ethical risk inherent in this technology and introduces a novel method, termed the Cognitive Morphing Attack (CogMorph), which manipulates T2I models to generate images that retain the original core subjects but embeds toxic or harmful contextual elements. This nuanced manipulation exploits the cognitive principle that human perception of concepts is shaped by the entire visual scene and its context, producing images that amplify emotional harm far beyond attacks that merely preserve the original semantics. To address this, we first construct an imagery toxicity taxonomy spanning 10 major and 48 sub-categories, aligned with human cognitive-perceptual dimensions, and further build a toxicity risk matrix resulting in 1,176 high-quality T2I toxic prompts. Based on this, our CogMorph first introduces Cognitive Toxicity Augmentation, which develops a cognitive toxicity knowledge base with rich external toxic representations for humans (e.g., fine-grained visual features) that can be utilized to further guide the optimization of adversarial prompts. In addition, we present Contextual Hierarchical Morphing, which hierarchically extracts critical parts of the original prompt (e.g., scenes, subjects, and body parts), and then iteratively retrieves and fuses toxic features to inject harmful contexts. Extensive experiments on multiple open-sourced T2I models and black-box commercial APIs (e.g., DALLE-3) demonstrate the efficacy of CogMorph which significantly outperforms other baselines by large margins (+20.62% on average).

Zonghao Ying, Siyang Wu, Run Hao et al.

Multimodal Large Language Models (MLLMs) have enabled transformative advancements across diverse applications but remain susceptible to safety threats, especially jailbreak attacks that induce harmful outputs. To systematically evaluate and improve their safety, we organized the Adversarial Testing & Large-model Alignment Safety Grand Challenge (ATLAS) 2025}. This technical report presents findings from the competition, which involved 86 teams testing MLLM vulnerabilities via adversarial image-text attacks in two phases: white-box and black-box evaluations. The competition results highlight ongoing challenges in securing MLLMs and provide valuable guidance for developing stronger defense mechanisms. The challenge establishes new benchmarks for MLLM safety evaluation and lays groundwork for advancing safer multimodal AI systems. The code and data for this challenge are openly available at https://github.com/NY1024/ATLAS_Challenge_2025.

Ge Yang, Changyi He, Jinyang Guo et al.

Although large language models (LLMs) have demonstrated their strong intelligence ability, the high demand for computation and storage hinders their practical application. To this end, many model compression techniques are proposed to increase the efficiency of LLMs. However, current researches only validate their methods on limited models, datasets, metrics, etc, and still lack a comprehensive evaluation under more general scenarios. So it is still a question of which model compression approach we should use under a specific case. To mitigate this gap, we present the Large Language Model Compression Benchmark (LLMCBench), a rigorously designed benchmark with an in-depth analysis for LLM compression algorithms. We first analyze the actual model production requirements and carefully design evaluation tracks and metrics. Then, we conduct extensive experiments and comparison using multiple mainstream LLM compression approaches. Finally, we perform an in-depth analysis based on the evaluation and provide useful insight for LLM compression design. We hope our LLMCBench can contribute insightful suggestions for LLM compression algorithm design and serve as a foundation for future research. Our code is available at https://github.com/AboveParadise/LLMCBench.

Le Wang, Zonghao Ying, Xiao Yang et al.

Embodied agents powered by vision-language models (VLMs) are increasingly capable of executing complex real-world tasks, yet they remain vulnerable to hazardous instructions that may trigger unsafe behaviors. Runtime safety guardrails, which intercept hazardous actions during task execution, offer a promising solution due to their flexibility. However, existing defenses often rely on static rule filters or prompt-level control, which struggle to address implicit risks arising in dynamic, temporally dependent, and context-rich environments. To address this, we propose RoboSafe, a hybrid reasoning runtime safeguard for embodied agents through executable predicate-based safety logic. RoboSafe integrates two complementary reasoning processes on a Hybrid Long-Short Safety Memory. We first propose a Backward Reflective Reasoning module that continuously revisits recent trajectories in short-term memory to infer temporal safety predicates and proactively triggers replanning when violations are detected. We then propose a Forward Predictive Reasoning module that anticipates upcoming risks by generating context-aware safety predicates from the long-term safety memory and the agent's multimodal observations. Together, these components form an adaptive, verifiable safety logic that is both interpretable and executable as code. Extensive experiments across multiple agents demonstrate that RoboSafe substantially reduces hazardous actions (-36.8% risk occurrence) compared with leading baselines, while maintaining near-original task performance. Real-world evaluations on physical robotic arms further confirm its practicality. Code will be released upon acceptance.

Zining Wnag, Jinyang Guo, Ruihao Gong et al.

With the increased attention to model efficiency, post-training sparsity (PTS) has become more and more prevalent because of its effectiveness and efficiency. However, there remain questions on better practice of PTS algorithms and the sparsification ability of models, which hinders the further development of this area. Therefore, a benchmark to comprehensively investigate the issues above is urgently needed. In this paper, we propose the first comprehensive post-training sparsity benchmark called PTSBench towards algorithms and models. We benchmark 10+ PTS general-pluggable fine-grained techniques on 3 typical tasks using over 40 off-the-shelf model architectures. Through extensive experiments and analyses, we obtain valuable conclusions and provide several insights from both algorithms and model aspects. Our PTSBench can provide (1) new observations for a better understanding of the PTS algorithms, (2) in-depth and comprehensive evaluations for the sparsification ability of models, and (3) a well-structured and easy-integrate open-source framework. We hope this work will provide illuminating conclusions and advice for future studies of post-training sparsity methods and sparsification-friendly model design. The code for our PTSBench is released at \href{https://github.com/ModelTC/msbench}{https://github.com/ModelTC/msbench}.

Zonghao Ying, Haowen Dai, Tianyuan Zhang et al.

Self-evolving agents offer a promising path toward scalable autonomy. However, in this work, we show that in competitive environments, self-evolution can instead give rise to a serious and previously underexplored risk: the spontaneous emergence of deception as an evolutionarily stable strategy. We conduct a systematic empirical study on the self-evolution of large language model (LLM) agents in a competitive Bidding Arena, where agents iteratively refine their strategies through interaction-driven reflection. Across different evolutionary paths (\eg, Neutral, Honesty-Guided, and Deception-Guided), we find a consistent pattern: under utility-driven competition, unconstrained self-evolution reliably drifts toward deceptive behaviors, even when honest strategies remain viable. This drift is explained by a fundamental asymmetry in generalization. Deception evolves as a transferable meta-strategy that generalizes robustly across diverse and unseen tasks, whereas honesty-based strategies are fragile and often collapse outside their original contexts. Further analysis of agents internal states reveals the emergence of rationalization mechanisms, through which agents justify or deny deceptive actions to reconcile competitive success with normative instructions. Our paper exposes a fundamental tension between agent self-evolution and alignment, highlighting the risks of deploying self-improving agents in adversarial environments.

Jian Yang, Wei Zhang, Jiajun Wu et al.

Industrial software development across chip design, GPU optimization, and embedded systems lacks expert reasoning traces showing how engineers reason about hardware constraints and timing semantics. In this work, we propose InCoder-32B-Thinking, trained on the data from the Error-driven Chain-of-Thought (ECoT) synthesis framework with an industrial code world model (ICWM) to generate reasoning traces. Specifically, ECoT generates reasoning chains by synthesizing the thinking content from multi-turn dialogue with environmental error feedback, explicitly modeling the error-correction process. ICWM is trained on domain-specific execution traces from Verilog simulation, GPU profiling, etc., learns the causal dynamics of how code affects hardware behavior, and enables self-verification by predicting execution outcomes before actual compilation. All synthesized reasoning traces are validated through domain toolchains, creating training data matching the natural reasoning depth distribution of industrial tasks. Evaluation on 14 general (81.3% on LiveCodeBench v5) and 9 industrial benchmarks (84.0% in CAD-Coder and 38.0% on KernelBench) shows InCoder-32B-Thinking achieves top-tier open-source results across all domains.GPU Optimization

Jiayang Liu, Siyuan Liang, Shiqian Zhao et al.

In recent years, fueled by the rapid advancement of diffusion models, text-to-video (T2V) generation models have achieved remarkable progress, with notable examples including Pika, Luma, Kling, and Open-Sora. Although these models exhibit impressive generative capabilities, they also expose significant security risks due to their vulnerability to jailbreak attacks, where the models are manipulated to produce unsafe content such as pornography, violence, or discrimination. Existing works such as T2VSafetyBench provide preliminary benchmarks for safety evaluation, but lack systematic methods for thoroughly exploring model vulnerabilities. To address this gap, we are the first to formalize the T2V jailbreak attack as a discrete optimization problem and propose a joint objective-based optimization framework, called T2V-OptJail. This framework consists of two key optimization goals: bypassing the built-in safety filtering mechanisms to increase the attack success rate, preserving semantic consistency between the adversarial prompt and the unsafe input prompt, as well as between the generated video and the unsafe input prompt, to enhance content controllability. In addition, we introduce an iterative optimization strategy guided by prompt variants, where multiple semantically equivalent candidates are generated in each round, and their scores are aggregated to robustly guide the search toward optimal adversarial prompts. We conduct large-scale experiments on several T2V models, covering both open-source models and real commercial closed-source models. The experimental results show that the proposed method improves 11.4% and 10.0% over the existing state-of-the-art method in terms of attack success rate assessed by GPT-4, attack success rate assessed by human accessors, respectively, verifying the significant advantages of the method in terms of attack effectiveness and content control.

Yisong Xiao, Aishan Liu, Siyuan Liang et al.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, yet they remain vulnerable to generating toxic content, necessitating detoxification strategies to ensure safe and responsible deployment. Test-time detoxification methods, which typically introduce static or dynamic interventions into LLM representations, offer a promising solution due to their flexibility and minimal invasiveness. However, current approaches often suffer from imprecise interventions, primarily due to their insufficient exploration of the transition space between toxic and non-toxic outputs. To address this challenge, we propose \textsc{A}utoregressive \textsc{R}eward \textsc{G}uided \textsc{R}epresentation \textsc{E}diting (ARGRE), a novel test-time detoxification framework that explicitly models toxicity transitions within the latent representation space, enabling stable and precise reward-guided editing. ARGRE identifies non-toxic semantic directions and interpolates between toxic and non-toxic representations to reveal fine-grained transition trajectories. These trajectories transform sparse toxicity annotations into dense training signals, enabling the construction of an autoregressive reward model that delivers stable and precise editing guidance. At inference, the reward model guides an adaptive two-step editing process to obtain detoxified representations: it first performs directional steering based on expected reward gaps to shift representations toward non-toxic regions, followed by lightweight gradient-based refinements. Extensive experiments across 8 widely used LLMs show that ARGRE significantly outperforms leading baselines in effectiveness (-62.21% toxicity) and efficiency (-47.58% inference time), while preserving the core capabilities of the original model with minimal degradation. Our code is available at the website.

Shuhan Xu, Siyuan Liang, Hongling Zheng et al.

Visual language models (VLMs) have made significant progress in image captioning tasks, yet recent studies have found they are vulnerable to backdoor attacks. Attackers can inject undetectable perturbations into the data during inference, triggering abnormal behavior and generating malicious captions. These attacks are particularly challenging to detect and defend against due to the stealthiness and cross-modal propagation of the trigger signals. In this paper, we identify two key vulnerabilities by analyzing existing attack patterns: (1) the model exhibits abnormal attention concentration on certain regions of the input image, and (2) backdoor attacks often induce semantic drift and sentence incoherence. Based on these insights, we propose Semantic Reward Defense (SRD), a reinforcement learning framework that mitigates backdoor behavior without requiring any prior knowledge of trigger patterns. SRD learns to apply discrete perturbations to sensitive contextual regions of image inputs via a deep Q-network policy, aiming to confuse attention and disrupt the activation of malicious paths. To guide policy optimization, we design a reward signal named semantic fidelity score, which jointly assesses the semantic consistency and linguistic fluency of the generated captions, encouraging the agent to achieve a robust yet faithful output. SRD offers a trigger-agnostic, policy-interpretable defense paradigm that effectively mitigates local (TrojVLM) and global (Shadowcast) backdoor attacks, reducing ASR to 3.6% and 5.6% respectively, with less than 15% average CIDEr drop on the clean inputs. Our codes can be found at https://github.com/Ciconey/SRD.git.

Tianyuan Zhang, Peng Yue, Zihao Peng et al.

Multimodal large language models (MLLMs) are increasingly integrated into autonomous driving (AD) systems; however, they remain vulnerable to diverse safety threats, particularly in accident-prone scenarios. Recent safeguard mechanisms have shown promise by incorporating logical constraints, yet most rely on static formulations that lack temporally grounded safety reasoning over evolving traffic interactions, resulting in limited robustness in dynamic driving environments. To address these limitations, we propose GuardAD, a model-agnostic safeguard that formulates AD safety as an evolving Markovian logical state. GuardAD introduces Neuro-Symbolic Logic Formalization, which represents safety predicates over heterogeneous traffic participants and continuously induces them via n-th order Markovian Logic Induction. This design enables the inference of emerging and latent hazards beyond single-step observations. Rather than simply vetoing unsafe actions, GuardAD performs Logic-Driven Action Revision, where inferred safety states actively guide action refinement without modifying the underlying MLLM. Extensive experiments on multiple benchmarks and AD-MLLMs demonstrate that GuardAD substantially reduces accident rates (-32.07%) while slightly improving task performance (+6.85%). Moreover, closed-loop simulation evaluations, together with physical-world vehicle studies, further validate the effectiveness and potential of GuardAD.

Yisong Xiao, Aishan Liu, QianJia Cheng et al.

Large Vision-Language Models (LVLMs) have been widely adopted in various applications; however, they exhibit significant gender biases. Existing benchmarks primarily evaluate gender bias at the demographic group level, neglecting individual fairness, which emphasizes equal treatment of similar individuals. This research gap limits the detection of discriminatory behaviors, as individual fairness offers a more granular examination of biases that group fairness may overlook. For the first time, this paper introduces the GenderBias-\emph{VL} benchmark to evaluate occupation-related gender bias in LVLMs using counterfactual visual questions under individual fairness criteria. To construct this benchmark, we first utilize text-to-image diffusion models to generate occupation images and their gender counterfactuals. Subsequently, we generate corresponding textual occupation options by identifying stereotyped occupation pairs with high semantic similarity but opposite gender proportions in real-world statistics. This method enables the creation of large-scale visual question counterfactuals to expose biases in LVLMs, applicable in both multimodal and unimodal contexts through modifying gender attributes in specific modalities. Overall, our GenderBias-\emph{VL} benchmark comprises 34,581 visual question counterfactual pairs, covering 177 occupations. Using our benchmark, we extensively evaluate 15 commonly used open-source LVLMs (\eg, LLaVA) and state-of-the-art commercial APIs, including GPT-4o and Gemini-Pro. Our findings reveal widespread gender biases in existing LVLMs. Our benchmark offers: (1) a comprehensive dataset for occupation-related gender bias evaluation; (2) an up-to-date leaderboard on LVLM biases; and (3) a nuanced understanding of the biases presented by these models. \footnote{The dataset and code are available at the \href{https://genderbiasvl.github.io/}{website}.}

Zonghao Ying, Aishan Liu, Xianglong Liu et al.

The recent release of GPT-4o has garnered widespread attention due to its powerful general capabilities. While its impressive performance is widely acknowledged, its safety aspects have not been sufficiently explored. Given the potential societal impact of risky content generated by advanced generative AI such as GPT-4o, it is crucial to rigorously evaluate its safety. In response to this question, this paper for the first time conducts a rigorous evaluation of GPT-4o against jailbreak attacks. Specifically, this paper adopts a series of multi-modal and uni-modal jailbreak attacks on 4 commonly used benchmarks encompassing three modalities (ie, text, speech, and image), which involves the optimization of over 4,000 initial text queries and the analysis and statistical evaluation of nearly 8,000+ response on GPT-4o. Our extensive experiments reveal several novel observations: (1) In contrast to the previous version (such as GPT-4V), GPT-4o has enhanced safety in the context of text modality jailbreak; (2) The newly introduced audio modality opens up new attack vectors for jailbreak attacks on GPT-4o; (3) Existing black-box multimodal jailbreak attack methods are largely ineffective against GPT-4o and GPT-4V. These findings provide critical insights into the safety implications of GPT-4o and underscore the need for robust alignment guardrails in large models. Our code is available at \url{https://github.com/NY1024/Jailbreak_GPT4o}.

Xinwei Zhang, Aishan Liu, Tianyuan Zhang et al.

Deep learning-based lane detection (LD) plays a critical role in autonomous driving systems, such as adaptive cruise control. However, it is vulnerable to backdoor attacks. Existing backdoor attack methods on LD exhibit limited effectiveness in dynamic real-world scenarios, primarily because they fail to consider dynamic scene factors, including changes in driving perspectives (e.g., viewpoint transformations) and environmental conditions (e.g., weather or lighting changes). To tackle this issue, this paper introduces BadLANE, a dynamic scene adaptation backdoor attack for LD designed to withstand changes in real-world dynamic scene factors. To address the challenges posed by changing driving perspectives, we propose an amorphous trigger pattern composed of shapeless pixels. This trigger design allows the backdoor to be activated by various forms or shapes of mud spots or pollution on the road or lens, enabling adaptation to changes in vehicle observation viewpoints during driving. To mitigate the effects of environmental changes, we design a meta-learning framework to train meta-generators tailored to different environmental conditions. These generators produce meta-triggers that incorporate diverse environmental information, such as weather or lighting conditions, as the initialization of the trigger patterns for backdoor implantation, thus enabling adaptation to dynamic environments. Extensive experiments on various commonly used LD models in both digital and physical domains validate the effectiveness of our attacks, outperforming other baselines significantly (+25.15% on average in Attack Success Rate). Our codes will be available upon paper publication.

Aishan Liu, Xinwei Zhang, Yisong Xiao et al.

Pre-trained vision models (PVMs) have become a dominant component due to their exceptional performance when fine-tuned for downstream tasks. However, the presence of backdoors within PVMs poses significant threats. Unfortunately, existing studies primarily focus on backdooring PVMs for the classification task, neglecting potential inherited backdoors in downstream tasks such as detection and segmentation. In this paper, we propose the Pre-trained Trojan attack, which embeds backdoors into a PVM, enabling attacks across various downstream vision tasks. We highlight the challenges posed by cross-task activation and shortcut connections in successful backdoor attacks. To achieve effective trigger activation in diverse tasks, we stylize the backdoor trigger patterns with class-specific textures, enhancing the recognition of task-irrelevant low-level features associated with the target class in the trigger pattern. Moreover, we address the issue of shortcut connections by introducing a context-free learning pipeline for poison training. In this approach, triggers without contextual backgrounds are directly utilized as training data, diverging from the conventional use of clean images. Consequently, we establish a direct shortcut from the trigger to the target class, mitigating the shortcut connection issue. We conducted extensive experiments to thoroughly validate the effectiveness of our attacks on downstream detection and segmentation tasks. Additionally, we showcase the potential of our approach in more practical scenarios, including large vision models and 3D object detection in autonomous driving. This paper aims to raise awareness of the potential threats associated with applying PVMs in practical scenarios. Our codes will be available upon paper publication.

Simin Li, Shuing Zhang, Gujun Chen et al.

Physical world adversarial attack is a highly practical and threatening attack, which fools real world deep learning systems by generating conspicuous and maliciously crafted real world artifacts. In physical world attacks, evaluating naturalness is highly emphasized since human can easily detect and remove unnatural attacks. However, current studies evaluate naturalness in a case-by-case fashion, which suffers from errors, bias and inconsistencies. In this paper, we take the first step to benchmark and assess visual naturalness of physical world attacks, taking autonomous driving scenario as the first attempt. First, to benchmark attack naturalness, we contribute the first Physical Attack Naturalness (PAN) dataset with human rating and gaze. PAN verifies several insights for the first time: naturalness is (disparately) affected by contextual features (i.e., environmental and semantic variations) and correlates with behavioral feature (i.e., gaze signal). Second, to automatically assess attack naturalness that aligns with human ratings, we further introduce Dual Prior Alignment (DPA) network, which aims to embed human knowledge into model reasoning process. Specifically, DPA imitates human reasoning in naturalness assessment by rating prior alignment and mimics human gaze behavior by attentive prior alignment. We hope our work fosters researches to improve and automatically assess naturalness of physical world attacks. Our code and dataset can be found at https://github.com/zhangsn-19/PAN.

Shiyu Tang, Ruihao Gong, Yan Wang et al.

Deep neural networks (DNNs) are vulnerable to adversarial noises, which motivates the benchmark of model robustness. Existing benchmarks mainly focus on evaluating defenses, but there are no comprehensive studies of how architecture design and training techniques affect robustness. Comprehensively benchmarking their relationships is beneficial for better understanding and developing robust DNNs. Thus, we propose RobustART, the first comprehensive Robustness investigation benchmark on ImageNet regarding ARchitecture design (49 human-designed off-the-shelf architectures and 1200+ networks from neural architecture search) and Training techniques (10+ techniques, e.g., data augmentation) towards diverse noises (adversarial, natural, and system noises). Extensive experiments substantiated several insights for the first time, e.g., (1) adversarial training is effective for the robustness against all noises types for Transformers and MLP-Mixers; (2) given comparable model sizes and aligned training settings, CNNs > Transformers > MLP-Mixers on robustness against natural and system noises; Transformers > MLP-Mixers > CNNs on adversarial robustness; (3) for some light-weight architectures, increasing model sizes or using extra data cannot improve robustness. Our benchmark presents: (1) an open-source platform for comprehensive robustness evaluation; (2) a variety of pre-trained models to facilitate robustness evaluation; and (3) a new view to better understand the mechanism towards designing robust DNNs. We will continuously develop to this ecosystem for the community.