Selected work

Representative Papers

Deep Learning-based Intrusion Detection Systems: A Survey

Apr 10, 2025

This paper addresses the limited generalization capability of deep learning–based intrusion detection systems (DL-IDS) in detecting zero-day attacks. To tackle this challenge, it presents the first holistic, full-stack analysis of DL-IDS technical evolution—spanning data acquisition, log parsing, behavioral graph modeling, attack detection, and forensic traceability. The authors propose a unified framework integrating convolutional neural networks (CNNs), recurrent neural networks (RNNs), graph neural networks (GNNs), and self-supervised representation learning, enhanced by structured log parsing and dynamic graph summarization techniques. The survey systematically categorizes 12 mainstream methodologies, benchmarks performance across 7 publicly available datasets, and identifies 5 fundamental challenges. As the first comprehensive panorama dedicated to zero-day attack generalization in DL-IDS, this work establishes both theoretical foundations and practical guidelines for intelligent, adaptive cybersecurity detection.

1 citationsRead paper

WorldFly: A World-Model-Based Vision-Language-Action Model for UAV Navigation

Jun 04, 2026

This work addresses the challenge of robust end-to-end drone navigation in dense urban environments, where severe occlusions and abrupt viewpoint changes lead to incomplete observations and hinder reliable decision-making. To overcome this, the authors propose WorldFly, a novel framework that integrates a world model into a vision-language-action system for aerial navigation. WorldFly employs a dual-branch coupled flow-matching mechanism to jointly generate future video predictions and navigation actions, thereby enhancing spatial understanding and enabling imagination-driven policy planning. The study also introduces the Urban Canyon Traversal Benchmark for systematic evaluation. Experimental results demonstrate that WorldFly significantly outperforms existing baselines on this benchmark, exhibiting superior robustness and generalization—particularly in unseen urban scenarios.

0 citationsRead paper

TherapyProbe: Generating Design Knowledge for Relational Safety in Mental Health Chatbots Through Adversarial Simulation

Feb 26, 2026

This study addresses a critical gap in the safety evaluation of mental health chatbots, which has predominantly focused on single-turn crisis responses while neglecting relational risks that emerge over multi-turn interactions and may adversely affect users’ long-term well-being. To this end, the authors propose a reproducible, API-free adversarial multi-agent simulation framework that integrates dialogue trajectory analysis with clinical psychology theory to systematically identify 23 relational safety failure modes—such as “empathy fatigue” and “identity spirals.” Building upon these findings, they construct the first clinically grounded safety pattern library and translate it into actionable design guidelines tailored for developers, clinicians, and policymakers. This work substantially advances the capacity to understand, anticipate, and mitigate safety risks inherent in prolonged human–chatbot interactions within mental health contexts.

0 citationsRead paper

When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models

Feb 25, 2026

This study addresses the systematic erasure of cultural-linguistic markers from non-native English varieties by large language models (LLMs) during workplace text polishing, resulting in linguistic identity loss. Introducing the concept of “cultural ghosting,” the authors propose two novel metrics—Identity Erasure Rate (IER) and Semantic Preservation Score (SPS)—to quantify the extent of cultural erasure in Indian, Singaporean, and Nigerian English texts. Through large-scale generation experiments, cross-variety linguistic analysis, and cultural marker classification, the research reveals an average IER of 10.26%, with pragmatic markers being disproportionately removed compared to lexical ones. A culturally aware prompting strategy reduces erasure by 29% without compromising semantic fidelity, exposing a paradox wherein high semantic preservation coexists with significant cultural erasure.

0 citationsRead paper

Safe Urban Traffic Control via Uncertainty-Aware Conformal Prediction and World-Model Reinforcement Learning

Feb 04, 2026

This study addresses the critical need for reliable urban traffic management that jointly ensures accurate prediction, effective anomaly detection, and provably safe control. The authors propose STREAM-RL, a unified framework that, for the first time, propagates calibrated uncertainty end-to-end from prediction through anomaly detection to safety-aware policy learning, with formal theoretical guarantees. Key innovations include three novel algorithms: an uncertainty-guided graph attention network (PU-GAT+), a conformal residual flow network with Benjamini–Yekutieli false discovery rate (FDR) control (CRFN-BY), and a safe world model-based reinforcement learning method (LyCon-WRL+) equipped with Lyapunov stability certificates and Lipschitz bounds. Evaluated on real-world traffic data, the approach achieves 91.4% coverage efficiency, 4.1% FDR control, and a 95.2% safety rate—26.2 percentage points higher than PPO—while delivering higher rewards and maintaining an end-to-end inference latency of only 23 ms.

0 citationsRead paper

Recent publications

Latest Papers

WorldFly: A World-Model-Based Vision-Language-Action Model for UAV Navigation

Jun 04, 2026

This work addresses the challenge of robust end-to-end drone navigation in dense urban environments, where severe occlusions and abrupt viewpoint changes lead to incomplete observations and hinder reliable decision-making. To overcome this, the authors propose WorldFly, a novel framework that integrates a world model into a vision-language-action system for aerial navigation. WorldFly employs a dual-branch coupled flow-matching mechanism to jointly generate future video predictions and navigation actions, thereby enhancing spatial understanding and enabling imagination-driven policy planning. The study also introduces the Urban Canyon Traversal Benchmark for systematic evaluation. Experimental results demonstrate that WorldFly significantly outperforms existing baselines on this benchmark, exhibiting superior robustness and generalization—particularly in unseen urban scenarios.

0 citationsRead paper

TherapyProbe: Generating Design Knowledge for Relational Safety in Mental Health Chatbots Through Adversarial Simulation

Feb 26, 2026

This study addresses a critical gap in the safety evaluation of mental health chatbots, which has predominantly focused on single-turn crisis responses while neglecting relational risks that emerge over multi-turn interactions and may adversely affect users’ long-term well-being. To this end, the authors propose a reproducible, API-free adversarial multi-agent simulation framework that integrates dialogue trajectory analysis with clinical psychology theory to systematically identify 23 relational safety failure modes—such as “empathy fatigue” and “identity spirals.” Building upon these findings, they construct the first clinically grounded safety pattern library and translate it into actionable design guidelines tailored for developers, clinicians, and policymakers. This work substantially advances the capacity to understand, anticipate, and mitigate safety risks inherent in prolonged human–chatbot interactions within mental health contexts.

0 citationsRead paper

When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models

Feb 25, 2026

This study addresses the systematic erasure of cultural-linguistic markers from non-native English varieties by large language models (LLMs) during workplace text polishing, resulting in linguistic identity loss. Introducing the concept of “cultural ghosting,” the authors propose two novel metrics—Identity Erasure Rate (IER) and Semantic Preservation Score (SPS)—to quantify the extent of cultural erasure in Indian, Singaporean, and Nigerian English texts. Through large-scale generation experiments, cross-variety linguistic analysis, and cultural marker classification, the research reveals an average IER of 10.26%, with pragmatic markers being disproportionately removed compared to lexical ones. A culturally aware prompting strategy reduces erasure by 29% without compromising semantic fidelity, exposing a paradox wherein high semantic preservation coexists with significant cultural erasure.

0 citationsRead paper

Safe Urban Traffic Control via Uncertainty-Aware Conformal Prediction and World-Model Reinforcement Learning

Feb 04, 2026

This study addresses the critical need for reliable urban traffic management that jointly ensures accurate prediction, effective anomaly detection, and provably safe control. The authors propose STREAM-RL, a unified framework that, for the first time, propagates calibrated uncertainty end-to-end from prediction through anomaly detection to safety-aware policy learning, with formal theoretical guarantees. Key innovations include three novel algorithms: an uncertainty-guided graph attention network (PU-GAT+), a conformal residual flow network with Benjamini–Yekutieli false discovery rate (FDR) control (CRFN-BY), and a safe world model-based reinforcement learning method (LyCon-WRL+) equipped with Lyapunov stability certificates and Lipschitz bounds. Evaluated on real-world traffic data, the approach achieves 91.4% coverage efficiency, 4.1% FDR control, and a 95.2% safety rate—26.2 percentage points higher than PPO—while delivering higher rewards and maintaining an end-to-end inference latency of only 23 ms.

0 citationsRead paper

Cryptanalysis of Pseudorandom Error-Correcting Codes

Dec 19, 2025

This work addresses the unresolved security of pseudorandom error-correcting codes (PRCs)—a novel cryptographic primitive for AI watermarking—by conducting the first systematic cryptanalysis. We propose three practical attacks: (1) statistical distinguishability combined with algebraic decoding to break undetectability; (2) robustness-breaking attacks leveraging large-model output feature modeling; and (3) weak-key–driven key-recovery analysis. Evaluated on real-world models—including DeepSeek and Stable Diffusion—our attacks universally compromise PRC security guarantees across all tested parameters, achieving near-100% watermark detection rates. We further design an efficient detection attack with time complexity 2²² and propose a revised key-generation scheme resistant to weak-key exploitation. Our results demonstrate that existing PRC-based watermarking schemes fail to meet the 128-bit security requirement, revealing critical cryptographic vulnerabilities in current AI watermarking primitives.

0 citationsRead paper