arXiv subjects

Wenyun Li

Publications and source records attributed to Wenyun Li.

12 recordsLinked to original sources

arXiv2026

Text-Anchored Semantic Perturbations for Transferable Jailbreak Attacks on Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language interaction, yet their safety alignment remains vulnerable to jailbreak attacks. A key challenge is that safety behavior learned in the textual space does not reliably transfer to fused cross-modal representations, leaving multimodal inputs exploitable through latent semantic cues. We propose Text-Anchored Semantic Perturbation Attack (TA-SPA), a black-box jailbreak framework that optimizes transferable perturbations in a text-anchored semantic space. TA-SPA integrates Text-Anchored Semantic Factorization (TASF), which encourages the separation of cross-modal semantic factors from modality-specific residuals, with Semantic-Preserving Augmentation (SPA), which diversifies harmful target anchors while preserving semantic consistency. Experiments show strong attack effectiveness and transfer to commercial MLLMs, with competitive performance under representative defenses. Additional controls and probing support the intended factorization without implying perfect disentanglement, motivating representation-level safety alignment beyond input-level filtering.

cs.CL↗

arXiv2026

Shaping Sparse Rewards in Reinforcement Learning: A Semi-supervised Approach

In many real-world scenarios, reward signal for agents are exceedingly sparse, making it challenging to learn an effective reward function for reward shaping. To address this issue, the proposed approach in this paper performs reward shaping not only by utilizing non-zero-reward transitions but also by employing the \emph{Semi-Supervised Learning} (SSL) technique combined with a novel data augmentation to learn trajectory space representations from the majority of transitions, {i.e}., zero-reward transitions, thereby improving the efficacy of reward shaping. Experimental results in Atari and robotic manipulation demonstrate that our method outperforms supervised-based approaches in reward inference, leading to higher agent scores. Notably, in more sparse-reward environments, our method achieves up to twice the peak scores compared to supervised baselines. The proposed double entropy data augmentation enhances performance, showcasing a 15.8\% increase in best score over other augmentation methods

cs.LG↗

arXiv2026

Efficient Adversarial Training via Criticality-Aware Fine-Tuning

Vision Transformer (ViT) models have achieved remarkable performance across various vision tasks, with scalability being a key advantage when applied to large datasets. This scalability enables ViT models to exhibit strong generalization capabilities. However, as the number of parameters increases, the robustness of ViT models to adversarial examples does not scale proportionally. Adversarial training (AT), one of the most effective methods for enhancing robustness, typically requires fine-tuning the entire model, leading to prohibitively high computational costs, especially for large ViT architectures. In this paper, we aim to robustly fine-tune only a small subset of parameters to achieve robustness comparable to standard AT. To accomplish this, we introduce Criticality-Aware Adversarial Training (CAAT), a novel method that adaptively allocates resources to the most robustness-critical parameters, fine-tuning only selected modules. Specifically, CAAT efficiently identifies parameters that contribute most to adversarial robustness. It then leverages parameter-efficient fine-tuning (PEFT) to robustly adjust weight matrices where the number of critical parameters exceeds a predefined threshold. CAAT exhibits favorable generalization when scaled to larger vision transformer architectures, potentially paving the way for adversarial training at scale, e.g, compared with plain adversarial training, CAAT incurs only a 4.3% decrease in adversarial robustness while tuning approximately 6% of its parameters. Extensive experiments on three widely used adversarial learning datasets demonstrate that CAAT outperforms state-of-the-art lightweight AT methods with fewer trainable parameters.

cs.CV↗

arXiv2026

Map-Agnostic And Interactive Safety-Critical Scenario Generation via Multi-Objective Tree Search

Generating safety-critical scenarios is essential for validating the robustness of autonomous driving systems, yet existing methods often struggle to produce collisions that are both realistic and diverse while ensuring explicit interaction logic among traffic participants. This paper presents a novel framework for traffic-flow level safety-critical scenario generation via multi-objective Monte Carlo Tree Search (MCTS). We reframe trajectory feasibility and naturalistic behavior as optimization objectives within a unified evaluation function, enabling the discovery of diverse collision events without compromising realism. A hybrid Upper Confidence Bound (UCB) and Lower Confidence Bound (LCB) search strategy is introduced to balance exploratory efficiency with risk-averse decision-making. Furthermore, our method is map-agnostic and supports interactive scenario generation with each vehicle individually powered by SUMO's microscopic traffic models, enabling realistic agent behaviors in arbitrary geographic locations imported from OpenStreetMap. We validate our approach across four high-risk accident zones in Hong Kong's complex urban environments. Experimental results demonstrate that our framework achieves an 85\% collision failure rate while generating trajectories with superior feasibility and comfort metrics. The resulting scenarios exhibit greater complexity, as evidenced by increased vehicle mileage and CO$_2$ emissions. Our work provides a principled solution for stress testing autonomous vehicles through the generation of realistic yet infrequent corner cases at traffic-flow level.

cs.RO↗

arXiv2026

DRL-ORA: Distributional Reinforcement Learning with Online Risk Adaption

One of the main challenges in reinforcement learning (RL) is that the agent has to make decisions that would influence the future performance without having complete knowledge of the environment. Dynamically adjusting the level of epistemic risk during the learning process can help to achieve reliable policies in safety-critical settings with better efficiency. In this work, we propose a new framework, Distributional RL with Online Risk Adaptation (DRL-ORA). This framework quantifies both epistemic and implicit aleatory uncertainties in a unified manner and dynamically adjusts the epistemic risk levels by solving a total variation minimization problem online. The framework unifies the existing variants of risk adaption approaches and offers better explainability and flexibility. The selection of risk levels is performed efficiently via a grid search using a Follow-The-Leader-type algorithm, where the offline oracle also corresponds to a ''satisficing measure'' under a specially modified loss function. We show that DRL-ORA outperforms existing methods that rely on fixed risk levels or manually designed risk level adaptation in multiple classes of tasks.

cs.LG↗

arXiv2026

Generalizable Trajectory Prediction via Inverse Reinforcement Learning with Mamba-Graph Architecture

Accurate driving behavior modeling is fundamental to safe and efficient trajectory prediction, yet remains challenging in complex traffic scenarios. This paper presents a novel Inverse Reinforcement Learning (IRL) framework that captures human-like decision-making by inferring diverse reward functions, enabling robust cross-scenario adaptability. The learned reward function is utilized to maximize the likelihood of output by integrating Mamba blocks for efficient long-sequence dependency modeling with graph attention networks to encode spatial interactions among traffic agents. Comprehensive evaluations on urban intersections and roundabouts demonstrate that the proposed method not only outperforms various popular approaches in terms of prediction accuracy but also achieves 2.3 times higher generalization performance to unseen scenarios compared to other baselines, achieving adaptability in Out-of-Distribution settings that is competitive with fine-tuning.

cs.LG↗

arXiv2025

Bolster Hallucination Detection via Prompt-Guided Data Augmentation

Large language models (LLMs) have garnered significant interest in AI community. Despite their impressive generation capabilities, they have been found to produce misleading or fabricated information, a phenomenon known as hallucinations. Consequently, hallucination detection has become critical to ensure the reliability of LLM-generated content. One primary challenge in hallucination detection is the scarcity of well-labeled datasets containing both truthful and hallucinated outputs. To address this issue, we introduce Prompt-guided data Augmented haLlucination dEtection (PALE), a novel framework that leverages prompt-guided responses from LLMs as data augmentation for hallucination detection. This strategy can generate both truthful and hallucinated data under prompt guidance at a relatively low cost. To more effectively evaluate the truthfulness of the sparse intermediate embeddings produced by LLMs, we introduce an estimation metric called the Contrastive Mahalanobis Score (CM Score). This score is based on modeling the distributions of truthful and hallucinated data in the activation space. CM Score employs a matrix decomposition approach to more accurately capture the underlying structure of these distributions. Importantly, our framework does not require additional human annotations, offering strong generalizability and practicality for real-world applications. Extensive experiments demonstrate that PALE achieves superior hallucination detection performance, outperforming the competitive baseline by a significant margin of 6.55%.

cs.LG↗

arXiv2025

Transferable Adversarial Face Attack with Text Controlled Attribute

Traditional adversarial attacks typically produce adversarial examples under norm-constrained conditions, whereas unrestricted adversarial examples are free-form with semantically meaningful perturbations. Current unrestricted adversarial impersonation attacks exhibit limited control over adversarial face attributes and often suffer from low transferability. In this paper, we propose a novel Text Controlled Attribute Attack (TCA$^2$) to generate photorealistic adversarial impersonation faces guided by natural language. Specifically, the category-level personal softmax vector is employed to precisely guide the impersonation attacks. Additionally, we propose both data and model augmentation strategies to achieve transferable attacks on unknown target models. Finally, a generative model, \textit{i.e}, Style-GAN, is utilized to synthesize impersonated faces with desired attributes. Extensive experiments on two high-resolution face recognition datasets validate that our TCA$^2$ method can generate natural text-guided adversarial impersonation faces with high transferability. We also evaluate our method on real-world face recognition systems, \textit{i.e}, Face++ and Aliyun, further demonstrating the practical potential of our approach.

cs.CV↗

arXiv2023

ELF: An End-to-end Local and Global Multimodal Fusion Framework for Glaucoma Grading

Glaucoma is a chronic neurodegenerative condition that can lead to blindness. Early detection and curing are very important in stopping the disease from getting worse for glaucoma patients. The 2D fundus images and optical coherence tomography(OCT) are useful for ophthalmologists in diagnosing glaucoma. There are many methods based on the fundus images or 3D OCT volumes; however, the mining for multi-modality, including both fundus images and data, is less studied. In this work, we propose an end-to-end local and global multi-modal fusion framework for glaucoma grading, named ELF for short. ELF can fully utilize the complementary information between fundus and OCT. In addition, unlike previous methods that concatenate the multi-modal features together, which lack exploring the mutual information between different modalities, ELF can take advantage of local-wise and global-wise mutual information. The extensive experiment conducted on the multi-modal glaucoma grading GAMMA dataset can prove the effiectness of ELF when compared with other state-of-the-art methods.

eess.IV↗

arXiv2023

Locality Preserving Multiview Graph Hashing for Large Scale Remote Sensing Image Search

Hashing is very popular for remote sensing image search. This article proposes a multiview hashing with learnable parameters to retrieve the queried images for a large-scale remote sensing dataset. Existing methods always neglect that real-world remote sensing data lies on a low-dimensional manifold embedded in high-dimensional ambient space. Unlike previous methods, this article proposes to learn the consensus compact codes in a view-specific low-dimensional subspace. Furthermore, we have added a hyperparameter learnable module to avoid complex parameter tuning. In order to prove the effectiveness of our method, we carried out experiments on three widely used remote sensing data sets and compared them with seven state-of-the-art methods. Extensive experiments show that the proposed method can achieve competitive results compared to the other method.

cs.CV↗

arXiv2022

Asymmetric Scalable Cross-modal Hashing

Cross-modal hashing is a successful method to solve large-scale multimedia retrieval issue. A lot of matrix factorization-based hashing methods are proposed. However, the existing methods still struggle with a few problems, such as how to generate the binary codes efficiently rather than directly relax them to continuity. In addition, most of the existing methods choose to use an $n\times n$ similarity matrix for optimization, which makes the memory and computation unaffordable. In this paper we propose a novel Asymmetric Scalable Cross-Modal Hashing (ASCMH) to address these issues. It firstly introduces a collective matrix factorization to learn a common latent space from the kernelized features of different modalities, and then transforms the similarity matrix optimization to a distance-distance difference problem minimization with the help of semantic labels and common latent space. Hence, the computational complexity of the $n\times n$ asymmetric optimization is relieved. In the generation of hash codes we also employ an orthogonal constraint of label information, which is indispensable for search accuracy. So the redundancy of computation can be much reduced. For efficient optimization and scalable to large-scale datasets, we adopt the two-step approach rather than optimizing simultaneously. Extensive experiments on three benchmark datasets: Wiki, MIRFlickr-25K, and NUS-WIDE, demonstrate that our ASCMH outperforms the state-of-the-art cross-modal hashing methods in terms of accuracy and efficiency.

cs.CV↗

arXiv2021

A Single-Target License Plate Detection with Attention

With the development of deep learning, Neural Network is commonly adopted to the License Plate Detection (LPD) task and achieves much better performance and precision, especially CNN-based networks can achieve state of the art RetinaNet[1]. For a single object detection task such as LPD, modified general object detection would be time-consuming, unable to cope with complex scenarios and a cumbersome weights file that is too hard to deploy on the embedded device.

cs.CV↗