Selected work

Representative Papers

Interrupting the Chain: Human Perception of AI-Generated Disinformation Through a Kill Chain Lens

Jul 26, 2026

This study addresses the challenge that generative AI enables the scalable production of customized disinformation, while existing defenses remain largely reactive and lack a systematic understanding of the cognitive attack lifecycle. By adapting the cybersecurity kill chain model to the domain of cognitive security, this work conducts an empirical experiment with 504 participants to systematically analyze human perception and detection mechanisms regarding AI-generated disinformation. The research reveals a "perception-accuracy gap" and asymmetric cognitive fatigue effects, demonstrating that sustained exposure significantly reduces fake news detection rates by 10.2%. These findings precisely identify intervention points for proactive defense, providing critical theoretical and empirical foundations for constructing forward-looking, systematic cognitive security defense frameworks.

0 citationsRead paper

Good for the Planet, Bad for Me? Intended and Unintended Consequences of AI Energy Consumption Disclosure

Mar 24, 2026

This study addresses the trade-off users face between model performance and sustainability due to the high energy consumption of artificial intelligence systems. Through a randomized controlled experiment integrating behavioral measures, survey responses, and statistical modeling, we investigate how disclosing energy consumption information influences user preferences for small language models (SLMs). Results show that such disclosures increase the likelihood of selecting an SLM by more than twelvefold, yet do not significantly alter subsequent prompting behavior. Notably, users who chose SLMs reported lower satisfaction, revealing a pronounced negative perceptual bias. Our work identifies—for the first time—the “double-edged sword” effect of energy disclosure: while it encourages environmentally conscious choices, it concurrently undermines perceived model quality. These findings offer critical behavioral insights for designing sustainable AI systems.

0 citationsRead paper

Deep Generative and Discriminative Digital Twin endowed with Variational Autoencoder for Unsupervised Predictive Thermal Condition Monitoring of Physical Robots in Industry 6.0 and Society 6.0

Sep 16, 2025

In industrial and societal 6.0 contexts, robot motor overheating causes unplanned shutdowns, jeopardizing human–robot safety and system availability. To address this, we propose a generative-discriminative digital twin framework based on variational autoencoders (VAEs). Our method introduces the concept of “thermal difficulty,” modeling normal thermal behavior via unsupervised learning and quantifying motion-induced thermal risk through reconstruction error—enabling pre-execution thermal feasibility prediction. It operates without labeled data or physical cooling interventions, proactively avoiding hazardous motions and supporting cross-agent thermal knowledge sharing. Experiments demonstrate accurate early-warning capability prior to actual overheating, significantly enhancing equipment uptime and collaborative safety while overcoming the production continuity limitations imposed by conventional reactive shutdown mechanisms.

0 citationsRead paper

Deep Learning for Model-Free Prediction of Thermal States of Robot Joint Motors

Sep 16, 2025

Accurate prediction of joint motor thermal states in robotic systems is hindered by reliance on complex physics-based models, which suffer from difficult parameter acquisition, cumbersome calibration, and high uncertainty. Method: This paper proposes a fully data-driven, model-agnostic prediction framework that integrates Long Short-Term Memory (LSTM) networks with feedforward neural networks. It directly employs easily measurable time-series sensor data—such as joint torque—as input to learn temperature dynamics end-to-end, eliminating the need for physical modeling or parameter identification. Contribution/Results: The framework is highly scalable and requires no prior thermodynamic knowledge. Experiments on a 7-DOF redundant robotic platform demonstrate a prediction root-mean-square error below 0.8 °C, significantly outperforming conventional equivalent thermal circuit models and standalone feedforward networks. Results validate its high accuracy, robustness, and practical engineering applicability.

0 citationsRead paper

Sublinear Variational Optimization of Gaussian Mixture Models with Millions to Billions of Parameters

Jan 21, 2025

To address the prohibitive computational cost of training Gaussian Mixture Models (GMMs) on large-scale, high-dimensional data, this paper proposes a variational inference algorithm integrated with Mixture of Factor Analyzers (MFA). The method employs low-rank covariance parameterization and sublinear stochastic sampling, achieving—for the first time—per-iteration time complexity of *O(D)* in dimensionality *D* and *O(1)* in number of components *C*. Crucially, the total number of pairwise distance computations scales sublinearly with both sample size *N* and component count *C*. Empirically, the algorithm successfully trains a GMM with over 10 billion parameters on billion-scale image data, completing in approximately nine hours on a single CPU—more than ten times faster than the current state-of-the-art. This breakthrough significantly enhances the scalability and practical applicability of GMMs for ultra-large-scale problems.

0 citationsRead paper

Recent publications

Latest Papers

Interrupting the Chain: Human Perception of AI-Generated Disinformation Through a Kill Chain Lens

Jul 26, 2026

This study addresses the challenge that generative AI enables the scalable production of customized disinformation, while existing defenses remain largely reactive and lack a systematic understanding of the cognitive attack lifecycle. By adapting the cybersecurity kill chain model to the domain of cognitive security, this work conducts an empirical experiment with 504 participants to systematically analyze human perception and detection mechanisms regarding AI-generated disinformation. The research reveals a "perception-accuracy gap" and asymmetric cognitive fatigue effects, demonstrating that sustained exposure significantly reduces fake news detection rates by 10.2%. These findings precisely identify intervention points for proactive defense, providing critical theoretical and empirical foundations for constructing forward-looking, systematic cognitive security defense frameworks.

0 citationsRead paper

Good for the Planet, Bad for Me? Intended and Unintended Consequences of AI Energy Consumption Disclosure

Mar 24, 2026

This study addresses the trade-off users face between model performance and sustainability due to the high energy consumption of artificial intelligence systems. Through a randomized controlled experiment integrating behavioral measures, survey responses, and statistical modeling, we investigate how disclosing energy consumption information influences user preferences for small language models (SLMs). Results show that such disclosures increase the likelihood of selecting an SLM by more than twelvefold, yet do not significantly alter subsequent prompting behavior. Notably, users who chose SLMs reported lower satisfaction, revealing a pronounced negative perceptual bias. Our work identifies—for the first time—the “double-edged sword” effect of energy disclosure: while it encourages environmentally conscious choices, it concurrently undermines perceived model quality. These findings offer critical behavioral insights for designing sustainable AI systems.

0 citationsRead paper

Deep Generative and Discriminative Digital Twin endowed with Variational Autoencoder for Unsupervised Predictive Thermal Condition Monitoring of Physical Robots in Industry 6.0 and Society 6.0

Sep 16, 2025

In industrial and societal 6.0 contexts, robot motor overheating causes unplanned shutdowns, jeopardizing human–robot safety and system availability. To address this, we propose a generative-discriminative digital twin framework based on variational autoencoders (VAEs). Our method introduces the concept of “thermal difficulty,” modeling normal thermal behavior via unsupervised learning and quantifying motion-induced thermal risk through reconstruction error—enabling pre-execution thermal feasibility prediction. It operates without labeled data or physical cooling interventions, proactively avoiding hazardous motions and supporting cross-agent thermal knowledge sharing. Experiments demonstrate accurate early-warning capability prior to actual overheating, significantly enhancing equipment uptime and collaborative safety while overcoming the production continuity limitations imposed by conventional reactive shutdown mechanisms.

0 citationsRead paper

Deep Learning for Model-Free Prediction of Thermal States of Robot Joint Motors

Sep 16, 2025

Accurate prediction of joint motor thermal states in robotic systems is hindered by reliance on complex physics-based models, which suffer from difficult parameter acquisition, cumbersome calibration, and high uncertainty. Method: This paper proposes a fully data-driven, model-agnostic prediction framework that integrates Long Short-Term Memory (LSTM) networks with feedforward neural networks. It directly employs easily measurable time-series sensor data—such as joint torque—as input to learn temperature dynamics end-to-end, eliminating the need for physical modeling or parameter identification. Contribution/Results: The framework is highly scalable and requires no prior thermodynamic knowledge. Experiments on a 7-DOF redundant robotic platform demonstrate a prediction root-mean-square error below 0.8 °C, significantly outperforming conventional equivalent thermal circuit models and standalone feedforward networks. Results validate its high accuracy, robustness, and practical engineering applicability.

0 citationsRead paper

Sublinear Variational Optimization of Gaussian Mixture Models with Millions to Billions of Parameters

Jan 21, 2025

To address the prohibitive computational cost of training Gaussian Mixture Models (GMMs) on large-scale, high-dimensional data, this paper proposes a variational inference algorithm integrated with Mixture of Factor Analyzers (MFA). The method employs low-rank covariance parameterization and sublinear stochastic sampling, achieving—for the first time—per-iteration time complexity of *O(D)* in dimensionality *D* and *O(1)* in number of components *C*. Crucially, the total number of pairwise distance computations scales sublinearly with both sample size *N* and component count *C*. Empirically, the algorithm successfully trains a GMM with over 10 billion parameters on billion-scale image data, completing in approximately nine hours on a single CPU—more than ten times faster than the current state-of-the-art. This breakthrough significantly enhances the scalability and practical applicability of GMMs for ultra-large-scale problems.

0 citationsRead paper