Selected work

Representative Papers

Network- and Device-Level Cyber Deception for Contested Environments Using RL and LLMs

Mar 17, 2026

This work proposes a novel framework that integrates large language models (LLMs) with reinforcement learning (RL) to generate adaptive, intelligent deception strategies at both network and device levels—an approach not previously achieved. Addressing the limitations of traditional deception techniques, which are often costly, static, and reliant on manual intervention, the proposed method leverages containerization and operational technology (OT) security architectures to autonomously optimize and dynamically deploy deception mechanisms within simulated adversarial environments. Experimental results demonstrate that the framework significantly enhances defensive capabilities against stealthy attacks by improving deception efficacy while substantially reducing operational costs. The approach achieves high detection accuracy and superior cost-effectiveness, offering a scalable and intelligent solution for modern cyber defense.

0 citationsRead paper

A Conditional Diffusion Model for Building Energy Modeling Workflows

Nov 04, 2025

Urban energy modeling is often constrained by the scarcity of building-level attribute data, resulting in limited simulation fidelity and scalability. To address this, we propose the first conditional diffusion generative model tailored for building attribute imputation—innovatively adapting diffusion mechanisms to tabular building data, enabling conditional generation of mixed discrete and continuous features. Trained on a dataset of 2.2 million residential buildings, the model demonstrates high distributional fidelity in Baltimore: generated attributes closely match ground-truth distributions (Kolmogorov–Smirnov test p > 0.95). This substantially improves input completeness for energy simulations and enhances downstream prediction accuracy. Our approach establishes a scalable, high-fidelity data augmentation paradigm for large-scale urban energy modeling.

0 citationsRead paper

Multi-modal Bayesian Neural Network Surrogates with Conjugate Last-Layer Estimation

Sep 25, 2025

To address the low modeling efficiency and difficulty in uncertainty quantification for high-cost objective functions under multimodal data, this paper proposes two multimodal Bayesian neural network (BNN) surrogate models. The core methodological innovation is a conjugate final-layer design, enabling closed-form parameter updates and efficient variational inference while ensuring robustness to partial modality missing. By integrating multimodal feature encoding, stochastic variational inference, and conjugate distribution assumptions, the approach significantly improves prediction accuracy and uncertainty calibration on both scalar and time-series tasks—outperforming unimodal BNN baselines. The framework is modular and seamlessly embeddable into outer-loop applications such as optimization and inverse problem solving, thereby enhancing modeling efficiency, generalizability, and decision reliability for complex systems.

0 citationsRead paper

Summary Statistics of Large-scale Model Outputs for Observation-corrected Outputs

Jun 18, 2025

To address large biases in physics-based model outputs, sparse observational coverage, and massive data volumes, this paper proposes Sig-PCA—a spatiotemporal correction framework that operates on statistical summaries (rather than raw high-dimensional fields) of model output and fuses local observations via a lightweight neural network for efficient probabilistic calibration. Our method is the first to achieve observation-driven, dimensionality-reduced model correction while preserving the original spatiotemporal correlation structure and probabilistic characteristics. Innovatively, it employs Principal Component Analysis (PCA) to extract interpretable spatiotemporal principal components as statistical summaries and jointly models multi-source data. Evaluated on surface temperature and wind speed forecasting, Sig-PCA significantly improves accuracy in mean, variance, and spatiotemporal correlations; reduces probabilistic distribution calibration error by 32%; and decreases neural network parameter count by over 60%.

0 citationsRead paper

Leveraging Multi-Task Learning for Multi-Label Power System Security Assessment

May 09, 2025

This paper addresses the challenge of jointly modeling and interpreting multiple stability types—static, voltage, transient, and small-signal—in power system security assessment. To this end, we propose a multi-task learning (MTL)-based multi-label classification framework. Methodologically, we introduce, for the first time, a shared-encoder–multi-decoder neural architecture that unifies the four stability assessments into a single multi-label classification task, enabling both cross-task feature sharing and task-specific representation decoupling. Experimental results on the IEEE 68-bus system demonstrate that our approach achieves superior accuracy for all four stability classifications compared to state-of-the-art methods. The key contribution is establishing a novel paradigm for integrated multi-stability evaluation, delivering an end-to-end solution that simultaneously ensures high predictive accuracy and model interpretability for intelligent grid security analysis.

0 citationsRead paper

Recent publications

Latest Papers

Network- and Device-Level Cyber Deception for Contested Environments Using RL and LLMs

Mar 17, 2026

This work proposes a novel framework that integrates large language models (LLMs) with reinforcement learning (RL) to generate adaptive, intelligent deception strategies at both network and device levels—an approach not previously achieved. Addressing the limitations of traditional deception techniques, which are often costly, static, and reliant on manual intervention, the proposed method leverages containerization and operational technology (OT) security architectures to autonomously optimize and dynamically deploy deception mechanisms within simulated adversarial environments. Experimental results demonstrate that the framework significantly enhances defensive capabilities against stealthy attacks by improving deception efficacy while substantially reducing operational costs. The approach achieves high detection accuracy and superior cost-effectiveness, offering a scalable and intelligent solution for modern cyber defense.

0 citationsRead paper

A Conditional Diffusion Model for Building Energy Modeling Workflows

Nov 04, 2025

Urban energy modeling is often constrained by the scarcity of building-level attribute data, resulting in limited simulation fidelity and scalability. To address this, we propose the first conditional diffusion generative model tailored for building attribute imputation—innovatively adapting diffusion mechanisms to tabular building data, enabling conditional generation of mixed discrete and continuous features. Trained on a dataset of 2.2 million residential buildings, the model demonstrates high distributional fidelity in Baltimore: generated attributes closely match ground-truth distributions (Kolmogorov–Smirnov test p > 0.95). This substantially improves input completeness for energy simulations and enhances downstream prediction accuracy. Our approach establishes a scalable, high-fidelity data augmentation paradigm for large-scale urban energy modeling.

0 citationsRead paper

Multi-modal Bayesian Neural Network Surrogates with Conjugate Last-Layer Estimation

Sep 25, 2025

To address the low modeling efficiency and difficulty in uncertainty quantification for high-cost objective functions under multimodal data, this paper proposes two multimodal Bayesian neural network (BNN) surrogate models. The core methodological innovation is a conjugate final-layer design, enabling closed-form parameter updates and efficient variational inference while ensuring robustness to partial modality missing. By integrating multimodal feature encoding, stochastic variational inference, and conjugate distribution assumptions, the approach significantly improves prediction accuracy and uncertainty calibration on both scalar and time-series tasks—outperforming unimodal BNN baselines. The framework is modular and seamlessly embeddable into outer-loop applications such as optimization and inverse problem solving, thereby enhancing modeling efficiency, generalizability, and decision reliability for complex systems.

0 citationsRead paper

Summary Statistics of Large-scale Model Outputs for Observation-corrected Outputs

Jun 18, 2025

To address large biases in physics-based model outputs, sparse observational coverage, and massive data volumes, this paper proposes Sig-PCA—a spatiotemporal correction framework that operates on statistical summaries (rather than raw high-dimensional fields) of model output and fuses local observations via a lightweight neural network for efficient probabilistic calibration. Our method is the first to achieve observation-driven, dimensionality-reduced model correction while preserving the original spatiotemporal correlation structure and probabilistic characteristics. Innovatively, it employs Principal Component Analysis (PCA) to extract interpretable spatiotemporal principal components as statistical summaries and jointly models multi-source data. Evaluated on surface temperature and wind speed forecasting, Sig-PCA significantly improves accuracy in mean, variance, and spatiotemporal correlations; reduces probabilistic distribution calibration error by 32%; and decreases neural network parameter count by over 60%.

0 citationsRead paper

Leveraging Multi-Task Learning for Multi-Label Power System Security Assessment

May 09, 2025

This paper addresses the challenge of jointly modeling and interpreting multiple stability types—static, voltage, transient, and small-signal—in power system security assessment. To this end, we propose a multi-task learning (MTL)-based multi-label classification framework. Methodologically, we introduce, for the first time, a shared-encoder–multi-decoder neural architecture that unifies the four stability assessments into a single multi-label classification task, enabling both cross-task feature sharing and task-specific representation decoupling. Experimental results on the IEEE 68-bus system demonstrate that our approach achieves superior accuracy for all four stability classifications compared to state-of-the-art methods. The key contribution is establishing a novel paradigm for integrated multi-stability evaluation, delivering an end-to-end solution that simultaneously ensures high predictive accuracy and model interpretability for intelligent grid security analysis.

0 citationsRead paper