Selected work

Representative Papers

Shift Aware Transfer Learning with Adaptive Dual-Encoder Fusion for PM Forecasting in Data-Limited Environments

Aug 14, 2026

This study addresses the challenges of PM2.5 prediction and negative transfer arising from data scarcity and distribution shift by proposing a Distribution-Aware Adaptive Dual-Encoder Transfer Framework. The method adaptively integrates a pretrained source encoder with target-specific representations to preserve target information while effectively mitigating negative transfer. Validated through SHAP interpretability analysis and ablation studies, the model achieves an MSE of 21.66 and an R² of 0.8739, significantly outperforming baseline methods. These results demonstrate the framework’s efficacy in overcoming data limitations and statistical discrepancies, establishing a novel paradigm for cross-domain air quality forecasting.

0 citationsRead paper

A Deep Multiscale Neural Network for Accurate Neurological Disorder Detection from MRI Scans and Real-Time Web Deployment

Jun 27, 2026

This study addresses the challenge of capturing subtle anatomical differences in multi-class neurological disorder identification from MRI scans using deep convolutional neural networks, a task further complicated by class imbalance. To this end, the authors propose End-Net, a 24-layer multiscale deep network that integrates an enhanced Inception module—combining factorized 1×1, 3×3, and 5×5 convolutions with pooling—and a lightweight classification head. The architecture incorporates global average pooling and Dropout regularization, while leveraging WGAN-GP-based data augmentation and random undersampling to mitigate class imbalance. Evaluated on a four-class MRI dataset encompassing Alzheimer’s disease, brain tumors, multiple sclerosis, and healthy controls, End-Net significantly outperforms existing methods and enables end-to-end real-time deployment via a web interface for high-accuracy online inference.

0 citationsRead paper

Multi-Granular Discretization for Interpretable Generalization in Precise Cyberattack Identification

Jul 16, 2025

Existing interpretable intrusion detection systems (IDS) commonly rely on post-hoc approximators coupled with black-box classifiers, resulting in non-auditable rules, incomplete feature representation, and potential misinterpretation. To address this, we propose an end-to-end interpretable IDS framework: first, a multi-granularity Gaussian discretization method models continuous features in a human-readable, semantically meaningful manner; second, leveraging Interpretable Generalization, the framework directly learns auditable logical rules that distinguish benign from malicious traffic without approximation. The method achieves high accuracy and full transparency—even under extremely low training sample regimes—eliminating reliance on post-hoc surrogate models. Evaluated across nine distinct splits of the UKM-IDS20 dataset, it attains an average precision gain of ≥4 percentage points over state-of-the-art interpretable baselines, while maintaining near-perfect recall (≈1.0). These results demonstrate superior few-shot generalization capability and cross-dataset robustness.

0 citationsRead paper

GPU-Accelerated Interpretable Generalization for Rapid Cyberattack Detection and Forensics

Jul 16, 2025

Existing Interpretability-Guided (IG) mechanisms for intrusion detection offer high interpretability but suffer from cubic time complexity and prohibitively large intermediate bitset volumes, rendering them infeasible to scale to full datasets on CPU. This work presents the first end-to-end, sampling-free GPU acceleration of the entire IG pipeline—including dense intersection and subset operations—implemented natively in PyTorch. Evaluated on the full NSL-KDD dataset, our approach achieves training in 18 minutes (116× speedup over CPU) while attaining Recall = 0.957, Precision = 0.973, and AUC = 0.961; single-flow inference completes in milliseconds. Our core contribution is the first GPU-native IG architecture explicitly designed for interpretable pattern discovery—uniquely reconciling formal interpretability guarantees with industrial-scale scalability.

0 citationsRead paper

Recent publications

Latest Papers

Shift Aware Transfer Learning with Adaptive Dual-Encoder Fusion for PM Forecasting in Data-Limited Environments

Aug 14, 2026

This study addresses the challenges of PM2.5 prediction and negative transfer arising from data scarcity and distribution shift by proposing a Distribution-Aware Adaptive Dual-Encoder Transfer Framework. The method adaptively integrates a pretrained source encoder with target-specific representations to preserve target information while effectively mitigating negative transfer. Validated through SHAP interpretability analysis and ablation studies, the model achieves an MSE of 21.66 and an R² of 0.8739, significantly outperforming baseline methods. These results demonstrate the framework’s efficacy in overcoming data limitations and statistical discrepancies, establishing a novel paradigm for cross-domain air quality forecasting.

0 citationsRead paper

A Deep Multiscale Neural Network for Accurate Neurological Disorder Detection from MRI Scans and Real-Time Web Deployment

Jun 27, 2026

This study addresses the challenge of capturing subtle anatomical differences in multi-class neurological disorder identification from MRI scans using deep convolutional neural networks, a task further complicated by class imbalance. To this end, the authors propose End-Net, a 24-layer multiscale deep network that integrates an enhanced Inception module—combining factorized 1×1, 3×3, and 5×5 convolutions with pooling—and a lightweight classification head. The architecture incorporates global average pooling and Dropout regularization, while leveraging WGAN-GP-based data augmentation and random undersampling to mitigate class imbalance. Evaluated on a four-class MRI dataset encompassing Alzheimer’s disease, brain tumors, multiple sclerosis, and healthy controls, End-Net significantly outperforms existing methods and enables end-to-end real-time deployment via a web interface for high-accuracy online inference.

0 citationsRead paper

Multi-Granular Discretization for Interpretable Generalization in Precise Cyberattack Identification

Jul 16, 2025

Existing interpretable intrusion detection systems (IDS) commonly rely on post-hoc approximators coupled with black-box classifiers, resulting in non-auditable rules, incomplete feature representation, and potential misinterpretation. To address this, we propose an end-to-end interpretable IDS framework: first, a multi-granularity Gaussian discretization method models continuous features in a human-readable, semantically meaningful manner; second, leveraging Interpretable Generalization, the framework directly learns auditable logical rules that distinguish benign from malicious traffic without approximation. The method achieves high accuracy and full transparency—even under extremely low training sample regimes—eliminating reliance on post-hoc surrogate models. Evaluated across nine distinct splits of the UKM-IDS20 dataset, it attains an average precision gain of ≥4 percentage points over state-of-the-art interpretable baselines, while maintaining near-perfect recall (≈1.0). These results demonstrate superior few-shot generalization capability and cross-dataset robustness.

0 citationsRead paper

GPU-Accelerated Interpretable Generalization for Rapid Cyberattack Detection and Forensics

Jul 16, 2025

Existing Interpretability-Guided (IG) mechanisms for intrusion detection offer high interpretability but suffer from cubic time complexity and prohibitively large intermediate bitset volumes, rendering them infeasible to scale to full datasets on CPU. This work presents the first end-to-end, sampling-free GPU acceleration of the entire IG pipeline—including dense intersection and subset operations—implemented natively in PyTorch. Evaluated on the full NSL-KDD dataset, our approach achieves training in 18 minutes (116× speedup over CPU) while attaining Recall = 0.957, Precision = 0.973, and AUC = 0.961; single-flow inference completes in milliseconds. Our core contribution is the first GPU-native IG architecture explicitly designed for interpretable pattern discovery—uniquely reconciling formal interpretability guarantees with industrial-scale scalability.

0 citationsRead paper