Selected work

Representative Papers

Machine Learning-Based Cyber Defense for Cloud Infrastructure: An Adaptive Deep Q-Network Architecture for Intelligent Intrusion Detection and Automated Threat Mitigation

Aug 12, 2026

This study addresses the challenge of complex and evolving cyberattacks in cloud environments by proposing an adaptive dynamic defense framework based on reinforcement learning. The work introduces, for the first time, a Deep Q-Network (DQN) into cloud security to establish an end-to-end intelligent intrusion detection and automated response loop. Leveraging feature engineering and data preprocessing, the model is trained on the CICIDS2017 dataset and validated on UNSW-NB15. Experimental results demonstrate that the system achieves 99.72% accuracy, a 99.66% F1-score, and a 0.999 ROC-AUC under previously unseen and evolving attacks, with a false positive rate of only 0.31%, an average detection latency of 15 ms, and an attack mitigation rate of 99.54%, thereby significantly enhancing real-time performance and generalization capability.

0 citationsRead paper

Model-Informed Joint Material-Structural Optimization of Hard-Magnetic Soft Materials

Jul 15, 2026

This study addresses the challenges in predicting deformation and achieving co-optimized design of hard-magnetic soft materials under magnetic actuation by proposing a unified effective shear modulus framework. This framework integrates classical inclusion theory, the Hill self-consistent model, and constrained kinematic relations, and employs an experimentally calibrated Mooney strain energy function to formulate a multiphysics constitutive model. Building upon this foundation, a material–structure concurrent topology optimization method is developed to simultaneously tailor structural density, magnetic particle distribution, and remanent magnetization orientation. The proposed framework successfully generates non-intuitive designs capable of achieving prescribed deformations—such as rotation, translation, and recovery—demonstrating its versatility and precise controllability across single- and multi-loading scenarios and diverse design objectives.

0 citationsRead paper

Koopman Operator Identification of Model Parameter Trajectories for Temporal Domain Generalization (KOMET)

Mar 27, 2026

This work addresses the performance degradation of machine learning models in non-stationary environments caused by temporal domain drift by proposing a model-agnostic, zero-retraining adaptive framework. The approach models the sequence of model parameters as a trajectory of a nonlinear dynamical system and identifies its linear Koopman operator using extended dynamic mode decomposition (EDMD) with a Fourier-augmented observation dictionary. Leveraging a warm-start training protocol, the framework autonomously predicts future parameter trajectories without requiring future labels, enabling efficient adaptation. Moreover, it uncovers an interpretable dynamical structure underlying decision boundary drift. Evaluated across six datasets, the method achieves average accuracies between 0.981 and 1.000 over 100 future timesteps, demonstrating robustness and effectiveness under diverse distribution shift scenarios.

0 citationsRead paper

Content and Engagement Trends in COVID-19 YouTube Videos: Evidence from the Late Pandemic

Sep 02, 2025

This study investigates the mechanisms by which temporal distribution, title lexicon, content themes, and video duration influence user engagement with COVID-19–related videos on YouTube during the post-pandemic phase (2023–2024). Leveraging nearly 10,000 video metadata records and transcribed textual content, we apply time-series analysis, term-frequency statistics, sentiment analysis (validated via Pearson and Spearman correlation tests), and categorical descriptive modeling. Results reveal evolving audience behavior: engagement peaks midweek and on weekends; short-form videos—especially those explicitly labeled “Shorts”—achieve a mean view count of 2.16 million, significantly outperforming long-form content; and content genre moderates duration effects—person-centered and vlog-style long videos elicit higher interaction, whereas news- and politics-oriented videos consistently underperform in engagement metrics. This work provides the first systematic characterization of structured engagement patterns in post-pandemic digital health communication, offering empirically grounded insights for optimizing public health messaging in online platforms.

0 citationsRead paper

Systematic Classification of Studies Investigating Social Media Conversations about Long COVID Using a Novel Zero-Shot Transformer Framework

Mar 14, 2025

This study addresses the challenges of heterogeneous, manually annotated categorizations and topic ambiguity in Long COVID discussions on social media. We propose a zero-shot Transformer framework for metacategorization of medical literature—requiring neither labeled data nor predefined categories. Our approach integrates domain-adaptive prompt engineering, zero-shot text classification, and semantic similarity matching to automatically classify studies into four empirically grounded categories: clinical manifestations, computational methods, policy dissemination, and community support. Experimental results demonstrate an average classification confidence of 0.7788, significantly enhancing the efficiency, scalability, and generalizability of public health literature reviews. To our knowledge, this is the first zero-shot paradigm explicitly designed for medical literature metacategorization. It establishes a novel methodological pathway for large-scale analysis of health-related social discourse, bridging gaps between computational linguistics and public health research.

0 citationsRead paper

Recent publications

Latest Papers

Machine Learning-Based Cyber Defense for Cloud Infrastructure: An Adaptive Deep Q-Network Architecture for Intelligent Intrusion Detection and Automated Threat Mitigation

Aug 12, 2026

This study addresses the challenge of complex and evolving cyberattacks in cloud environments by proposing an adaptive dynamic defense framework based on reinforcement learning. The work introduces, for the first time, a Deep Q-Network (DQN) into cloud security to establish an end-to-end intelligent intrusion detection and automated response loop. Leveraging feature engineering and data preprocessing, the model is trained on the CICIDS2017 dataset and validated on UNSW-NB15. Experimental results demonstrate that the system achieves 99.72% accuracy, a 99.66% F1-score, and a 0.999 ROC-AUC under previously unseen and evolving attacks, with a false positive rate of only 0.31%, an average detection latency of 15 ms, and an attack mitigation rate of 99.54%, thereby significantly enhancing real-time performance and generalization capability.

0 citationsRead paper

Model-Informed Joint Material-Structural Optimization of Hard-Magnetic Soft Materials

Jul 15, 2026

This study addresses the challenges in predicting deformation and achieving co-optimized design of hard-magnetic soft materials under magnetic actuation by proposing a unified effective shear modulus framework. This framework integrates classical inclusion theory, the Hill self-consistent model, and constrained kinematic relations, and employs an experimentally calibrated Mooney strain energy function to formulate a multiphysics constitutive model. Building upon this foundation, a material–structure concurrent topology optimization method is developed to simultaneously tailor structural density, magnetic particle distribution, and remanent magnetization orientation. The proposed framework successfully generates non-intuitive designs capable of achieving prescribed deformations—such as rotation, translation, and recovery—demonstrating its versatility and precise controllability across single- and multi-loading scenarios and diverse design objectives.

0 citationsRead paper

Koopman Operator Identification of Model Parameter Trajectories for Temporal Domain Generalization (KOMET)

Mar 27, 2026

This work addresses the performance degradation of machine learning models in non-stationary environments caused by temporal domain drift by proposing a model-agnostic, zero-retraining adaptive framework. The approach models the sequence of model parameters as a trajectory of a nonlinear dynamical system and identifies its linear Koopman operator using extended dynamic mode decomposition (EDMD) with a Fourier-augmented observation dictionary. Leveraging a warm-start training protocol, the framework autonomously predicts future parameter trajectories without requiring future labels, enabling efficient adaptation. Moreover, it uncovers an interpretable dynamical structure underlying decision boundary drift. Evaluated across six datasets, the method achieves average accuracies between 0.981 and 1.000 over 100 future timesteps, demonstrating robustness and effectiveness under diverse distribution shift scenarios.

0 citationsRead paper

Content and Engagement Trends in COVID-19 YouTube Videos: Evidence from the Late Pandemic

Sep 02, 2025

This study investigates the mechanisms by which temporal distribution, title lexicon, content themes, and video duration influence user engagement with COVID-19–related videos on YouTube during the post-pandemic phase (2023–2024). Leveraging nearly 10,000 video metadata records and transcribed textual content, we apply time-series analysis, term-frequency statistics, sentiment analysis (validated via Pearson and Spearman correlation tests), and categorical descriptive modeling. Results reveal evolving audience behavior: engagement peaks midweek and on weekends; short-form videos—especially those explicitly labeled “Shorts”—achieve a mean view count of 2.16 million, significantly outperforming long-form content; and content genre moderates duration effects—person-centered and vlog-style long videos elicit higher interaction, whereas news- and politics-oriented videos consistently underperform in engagement metrics. This work provides the first systematic characterization of structured engagement patterns in post-pandemic digital health communication, offering empirically grounded insights for optimizing public health messaging in online platforms.

0 citationsRead paper

Systematic Classification of Studies Investigating Social Media Conversations about Long COVID Using a Novel Zero-Shot Transformer Framework

Mar 14, 2025

This study addresses the challenges of heterogeneous, manually annotated categorizations and topic ambiguity in Long COVID discussions on social media. We propose a zero-shot Transformer framework for metacategorization of medical literature—requiring neither labeled data nor predefined categories. Our approach integrates domain-adaptive prompt engineering, zero-shot text classification, and semantic similarity matching to automatically classify studies into four empirically grounded categories: clinical manifestations, computational methods, policy dissemination, and community support. Experimental results demonstrate an average classification confidence of 0.7788, significantly enhancing the efficiency, scalability, and generalizability of public health literature reviews. To our knowledge, this is the first zero-shot paradigm explicitly designed for medical literature metacategorization. It establishes a novel methodological pathway for large-scale analysis of health-related social discourse, bridging gaps between computational linguistics and public health research.

0 citationsRead paper