Selected work

Representative Papers

D3-Guard: Acoustic-based Drowsy Driving Detection Using Smartphones

Apr 01, 2019IEEE Conference on Computer Communications

This study addresses the limitation of existing fatigue-driving detection systems that rely on dedicated hardware. We propose a passive, smartphone-based acoustic sensing method leveraging built-in microphones to capture subtle Doppler-induced frequency shifts in ambient sound—caused by drowsiness-related behaviors such as yawning, head nodding, and steering wheel rotation. To enable efficient on-device processing, we introduce a lightweight undersampling–FFT feature extraction pipeline and develop an LSTM-based temporal model for early fatigue onset detection, achieving >80% detection accuracy within 70% of the behavioral event duration. To our knowledge, this is the first purely smartphone-microphone-driven acoustic fatigue detection framework. Evaluated on real-road driving data from five participants, the system achieves a mean classification accuracy of 93.31%, with low latency and strong potential for real-time, practical deployment.

42 citations4 influentialRead paper

Range and Topology Mutation Based Wireless Agility

Nov 09, 2020MTD@CCS

In hostile military environments, static wireless infrastructure is vulnerable to targeted denial-of-service (DoS) and eavesdropping attacks. To address this, we propose an active defense framework grounded in wireless agility. Our core method formally models two novel stochastic mutation primitives—Random Range Mutation (RNM) and Random Topology Mutation (RTM)—and jointly optimizes coverage continuity, security isolation, and energy constraints under unknown adversary strategies. We employ a hybrid Satisfiability Modulo Theories (SMT) and Answer Set Programming (ASP) solver to enforce dynamic client access-point (AP) handoffs while satisfying service requirements, thereby significantly enhancing resistance to device localization and targeted attacks. Experimental evaluation demonstrates the framework’s feasibility, scalability, and effectiveness in suppressing such attacks across large-scale deployments—marking a departure from conventional static defense paradigms.

6 citations1 influentialRead paper

SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models

Feb 05, 2025

Existing crystal generation methods struggle to simultaneously ensure crystallographic realism and structural novelty—either neglecting symmetry constraints entirely or mechanically reusing space-group information from databases. This work introduces the first diffusion-based framework that explicitly models and strictly preserves crystallographic symmetry. We propose a decoupled representation: an asymmetric unit coupled with generalizable space-group transformations. Furthermore, we design a group-action-driven geometric representation learning mechanism that jointly models atomic coordinates and symmetry operations throughout the diffusion process. Our method enables cross-space-group generalization and interpretable symmetry-aware generation. Evaluated on a subset of the Materials Project, it achieves 100% structural validity, high structural diversity, and exact space-group fidelity. Predicted material properties obey fundamental physical principles. Overall, our approach surpasses current state-of-the-art methods in both fidelity and generative capability.

2 citationsRead paper

Are Multimodal LLMs Ready for Surveillance? A Reality Check on Zero-Shot Anomaly Detection in the Wild

Mar 05, 2026

This work addresses the low zero-shot recall of multimodal large language models (MLLMs) in open-world video anomaly detection by reframing the task as binary classification under weak temporal supervision. The authors propose a language-guided reasoning framework and systematically evaluate MLLM performance on the ShanghaiTech and CHAD benchmarks. They reveal, for the first time, a conservative decision bias inherent in MLLMs for zero-shot anomaly detection and demonstrate that category-specific instructions effectively recalibrate decision boundaries. Through optimized prompt design, recall capability is substantially enhanced: with 1–3 second video clips as input, the F1 score on ShanghaiTech improves from 0.09 to 0.64. Nevertheless, achieving high recall remains a critical challenge for real-world deployment.

1 citationsRead paper

GenXSS: an AI-Driven Framework for Automated Detection of XSS Attacks in WAFs

Apr 11, 2025

Traditional WAFs struggle to detect highly obfuscated XSS attacks due to their reliance on manually curated, static rule sets and inherent inability to adapt to novel evasion techniques. This paper introduces the first end-to-end XSS attack-defense closed-loop framework powered by large language models (LLMs)—specifically GPT-4o and Gemini Pro—that integrates in-context learning with dynamic payload validation to jointly enable malicious payload generation, black-box WAF evasion testing, and reverse synthesis of deployable protection rules. Its key innovation lies in automating the full “generate–test–defend” pipeline using LLMs without requiring white-box WAF access. Experimentally, the framework generated 264 syntactically valid XSS payloads, 83% of which were functionally effective; 80% successfully bypassed the OWASP ModSecurity rule set. Moreover, just 15 automatically synthesized rules achieved 86% interception coverage against known evasion variants.

1 citationsRead paper

Recent publications

Latest Papers

The Game of Marginal Utilities

Aug 10, 2026

This study addresses the non-cooperative game-theoretic equilibrium of multi-player resource allocation in competitive projects, where payoffs are jointly influenced by diminishing marginal returns and congestion effects. By invoking the equimarginal principle, the authors characterize the structure of Nash equilibria, establishing for the first time the existence and uniqueness of equilibrium in this class of games. They further reveal that players allocate resources according to project priority rankings, resulting in nested investment intervals. In the fully active regime, equilibrium computation reduces to solving a single-variable nonlinear equation. Building on these insights, the paper proposes a structure-aware Block Pandora algorithm that integrates projected marginal utility updates, achieving global linear convergence and enabling efficient, high-precision reconstruction of individual strategies and payoffs.

0 citationsRead paper

MAVISEG: Manifold Propagation and Visual Prototypes for Zero-Shot Open-Vocabulary Segmentation in Diffusion Transformers

Aug 06, 2026

Existing zero-shot open-vocabulary semantic segmentation methods fail to exploit structured signals—such as temporal generation trajectories, conceptual visual statistics, and geometric relationships in pixel-level features—when leveraging diffusion Transformers. To address this limitation, this work proposes MAVISEG, a training-free, general-purpose post-processing framework that, for the first time, integrates these three types of structured information into the zero-shot segmentation pipeline. MAVISEG reconstructs semantic signals discarded by attribution methods through manifold propagation and visual prototype mechanisms operating jointly in the pixel-concept score field and feature space, without relying on any specific attribution strategy. Experiments demonstrate that MAVISEG achieves state-of-the-art training-free performance across six benchmarks, consistently outperforming prior methods in mIoU, with the most pronounced gains observed when the initial attribution quality is poorest.

0 citationsRead paper

Domain-Grounded Candidate Selection for Agentic Image Editing: A Shadow Removal Case

Aug 06, 2026

This work addresses the tendency of existing commercial generative editors to produce physically implausible results in shadow removal by erroneously treating shadows as intrinsic material or structural attributes, often due to neglecting the physical principles of illumination. To mitigate this, the authors propose an agent-based iterative editing framework that integrates physical priors of shadow formation—specifically, that shadows arise from occlusion of light sources. The framework employs a generator to construct guided probes and an evaluator that leverages these physical priors to detect inconsistencies and trigger retries. By sampling multiple candidates and selecting the output with highest physical consistency, the method explicitly embeds shadow physics into both prompt engineering and evaluation. Evaluated on the ShadowRemovalRefine benchmark, it achieves a CDD score of 0.0075, a 47% improvement over the current state of the art, significantly enhancing both shadow removal quality and scene fidelity.

0 citationsRead paper

CrystalGRPO: Target-Aligned and Coverage-Preserving Reinforcement Learning for Flow-Based Crystal Structure Prediction

Aug 06, 2026

This work addresses limitations in existing flow-based crystal generation models, which fail to directly optimize target structure recovery during pretraining and rely solely on energy-based rewards and random coordinate perturbations in reinforcement learning fine-tuning, often compromising polymorph identification accuracy and structural diversity. To overcome these issues, the authors propose CrystalGRPO, a novel framework that extends the ODE-to-SDE strategy to the joint coordinate–lattice space for the first time. Integrating MACE-based energy prediction with StructureMatcher-derived recovery scores, CrystalGRPO introduces two variants: CrystalGRPO-Q, which optimizes single-sample recovery, and CrystalGRPO-C, which incorporates trajectory regularization and coverage-aware advantage estimation. Evaluated on MP-20 and MPTS-52 benchmarks, CrystalGRPO-Q significantly improves Top-1 recovery accuracy, while CrystalGRPO-C enhances Top-20 success rates and reduces RMSE, effectively balancing precision and diversity.

0 citationsRead paper

VQ-VAD: Vector-quantized Motion Representation Learning for Human-centric Video Anomaly Detection

Aug 05, 2026

This work addresses the challenge of video anomaly detection when trained solely on normal data, where visual noise, appearance variations, and privacy constraints hinder robust modeling of human motion. To overcome these limitations, the authors propose an unsupervised detection framework based on a discrete motion codebook. For the first time, vector quantization is introduced into pose sequence modeling by adapting VQ-GAN to temporal keypoint data, mapping human actions into a discrete codebook space. Anomalies are identified through reconstruction errors in this discretized representation. By circumventing the constraints of continuous latent spaces on compact motion encoding, the method achieves 81.83% in-domain accuracy on HR-SHT and demonstrates strong generalization, attaining 76.69% cross-domain performance without fine-tuning.

0 citationsRead paper