Selected work

Representative Papers

Data Safety: Synthetic Data Quality Analysis Using CIFAKE Dataset

Jul 13, 2026

This work addresses the critical gap in understanding whether synthetic images are truly interchangeable with real ones in model training and the absence of systematic evaluation frameworks to ensure their safe and effective use. The study systematically quantifies discrepancies between synthetic and real images across three dimensions: high-dimensional feature distributions, low-level statistical properties in color space, and model training dynamics. Building on these insights, the authors propose a pre-evaluation metric for synthetic data of unknown quality and a safety-aware data fusion strategy for training. Experiments demonstrate that carefully calibrated mixing ratios and integration methods of synthetic and real data can substantially enhance model performance and robustness, thereby offering both theoretical grounding and practical guidance for the reliable deployment of synthetic data in machine learning pipelines.

0 citationsRead paper

SHAP-Weighted Cross-Modal Expert Fusion for Emotion and Sentiment Recognition: Evidence and Limits

Jul 09, 2026

This work addresses the limitations of early fusion—lacking modularity—and late fusion—neglecting cross-modal interactions—in multimodal sentiment recognition by proposing xgaf, an adaptive fusion method grounded in TreeSHAP attribution. xgaf employs a tree-based mixture-of-experts architecture to dynamically weight unimodal and cross-modal experts. Through systematic evaluation of various SHAP reduction strategies, the study identifies sum-abs as particularly effective, as it preserves total attribution magnitude while enhancing performance. The primary performance gain stems from incorporating trimodal experts rather than from complex routing mechanisms. On the MELD and CMU-MOSEI datasets, xgaf achieves weighted F1 scores of 0.5983 and 0.6519, respectively—significantly outperforming late fusion and matching or slightly surpassing early fusion—while simultaneously maintaining modularity and effectively modeling cross-modal interactions.

0 citationsRead paper

Cyber-Physical Anomaly Detection in IoT-Enabled Smart Grids Using Machine Learning and Metaheuristic Feature Optimization

May 21, 2026

This study addresses the challenge of distinguishing between physical faults and false data injection attacks in IoT-enabled smart grids, which often manifest as indistinguishable anomalies. To tackle this issue, the authors propose a lightweight detection framework that integrates genetic algorithms with tree-based ensemble models—specifically Extra Trees, XGBoost, and Random Forest—to enable metaheuristic-driven feature selection for dimensionality reduction of PMU/IED measurements. Evaluated on the MSU/ORNL power system attack dataset, the proposed GA+Extra Trees approach reduces the feature dimensionality from 112 to an average of 27.4 while achieving a macro-F1 score of 0.9212 and a ROC-AUC of 0.9837. The results demonstrate that the method not only significantly mitigates feature redundancy but also enhances both detection performance and model interpretability.

0 citationsRead paper

GPU-Accelerated Deep Learning for Heatwave Prediction and Urban Heat Risk Assessment

May 14, 2026

This study addresses the growing threat of urban heatwaves, exacerbated by climate change, by proposing a GPU-accelerated ConvLSTM deep learning framework for accurate next-day thermal environment forecasting and integrated heat risk assessment. The model uniquely combines MODIS land surface temperature data with Open-Meteo meteorological forecasts and introduces mixed-precision training—a first in urban heatwave prediction—to enhance computational efficiency. By incorporating exposure and vulnerability indicators, the framework enables multidimensional heat risk evaluation. Evaluated in Sarajevo, the model achieves high predictive accuracy with an MAE of 0.2293, RMSE of 0.3089, and R² of 0.8877, while significantly reducing training time and generating high-resolution heat risk maps for actionable urban planning insights.

0 citationsRead paper

Exploring the Impact of Skin Color on Skin Lesion Segmentation

Mar 31, 2026

This study addresses the limitations of existing skin lesion segmentation approaches that rely on coarse-grained, discrete skin tone categories, which obscure the true sources of performance disparities. On the HAM10000 and ISIC2017 datasets, the authors employ models including UNet, DeepLabV3 (with ResNet50 backbone), and DINOv2 to introduce, for the first time, a continuous pigmentation analysis based on pixel-level individual typology angle (ITA). They further quantify lesion-to-surrounding-skin contrast using Wasserstein distance. Their findings reveal that global skin lightness exhibits only a weak correlation with segmentation performance, whereas low lesion-skin contrast significantly increases segmentation error, indicating that boundary ambiguity and poor contrast are primary causes of model failure. This work thus offers a novel perspective and quantitative foundation for improving the fairness and robustness of segmentation algorithms across diverse skin tones.

0 citationsRead paper

Recent publications

Latest Papers

Data Safety: Synthetic Data Quality Analysis Using CIFAKE Dataset

Jul 13, 2026

This work addresses the critical gap in understanding whether synthetic images are truly interchangeable with real ones in model training and the absence of systematic evaluation frameworks to ensure their safe and effective use. The study systematically quantifies discrepancies between synthetic and real images across three dimensions: high-dimensional feature distributions, low-level statistical properties in color space, and model training dynamics. Building on these insights, the authors propose a pre-evaluation metric for synthetic data of unknown quality and a safety-aware data fusion strategy for training. Experiments demonstrate that carefully calibrated mixing ratios and integration methods of synthetic and real data can substantially enhance model performance and robustness, thereby offering both theoretical grounding and practical guidance for the reliable deployment of synthetic data in machine learning pipelines.

0 citationsRead paper

SHAP-Weighted Cross-Modal Expert Fusion for Emotion and Sentiment Recognition: Evidence and Limits

Jul 09, 2026

This work addresses the limitations of early fusion—lacking modularity—and late fusion—neglecting cross-modal interactions—in multimodal sentiment recognition by proposing xgaf, an adaptive fusion method grounded in TreeSHAP attribution. xgaf employs a tree-based mixture-of-experts architecture to dynamically weight unimodal and cross-modal experts. Through systematic evaluation of various SHAP reduction strategies, the study identifies sum-abs as particularly effective, as it preserves total attribution magnitude while enhancing performance. The primary performance gain stems from incorporating trimodal experts rather than from complex routing mechanisms. On the MELD and CMU-MOSEI datasets, xgaf achieves weighted F1 scores of 0.5983 and 0.6519, respectively—significantly outperforming late fusion and matching or slightly surpassing early fusion—while simultaneously maintaining modularity and effectively modeling cross-modal interactions.

0 citationsRead paper

Cyber-Physical Anomaly Detection in IoT-Enabled Smart Grids Using Machine Learning and Metaheuristic Feature Optimization

May 21, 2026

This study addresses the challenge of distinguishing between physical faults and false data injection attacks in IoT-enabled smart grids, which often manifest as indistinguishable anomalies. To tackle this issue, the authors propose a lightweight detection framework that integrates genetic algorithms with tree-based ensemble models—specifically Extra Trees, XGBoost, and Random Forest—to enable metaheuristic-driven feature selection for dimensionality reduction of PMU/IED measurements. Evaluated on the MSU/ORNL power system attack dataset, the proposed GA+Extra Trees approach reduces the feature dimensionality from 112 to an average of 27.4 while achieving a macro-F1 score of 0.9212 and a ROC-AUC of 0.9837. The results demonstrate that the method not only significantly mitigates feature redundancy but also enhances both detection performance and model interpretability.

0 citationsRead paper

GPU-Accelerated Deep Learning for Heatwave Prediction and Urban Heat Risk Assessment

May 14, 2026

This study addresses the growing threat of urban heatwaves, exacerbated by climate change, by proposing a GPU-accelerated ConvLSTM deep learning framework for accurate next-day thermal environment forecasting and integrated heat risk assessment. The model uniquely combines MODIS land surface temperature data with Open-Meteo meteorological forecasts and introduces mixed-precision training—a first in urban heatwave prediction—to enhance computational efficiency. By incorporating exposure and vulnerability indicators, the framework enables multidimensional heat risk evaluation. Evaluated in Sarajevo, the model achieves high predictive accuracy with an MAE of 0.2293, RMSE of 0.3089, and R² of 0.8877, while significantly reducing training time and generating high-resolution heat risk maps for actionable urban planning insights.

0 citationsRead paper

Exploring the Impact of Skin Color on Skin Lesion Segmentation

Mar 31, 2026

This study addresses the limitations of existing skin lesion segmentation approaches that rely on coarse-grained, discrete skin tone categories, which obscure the true sources of performance disparities. On the HAM10000 and ISIC2017 datasets, the authors employ models including UNet, DeepLabV3 (with ResNet50 backbone), and DINOv2 to introduce, for the first time, a continuous pigmentation analysis based on pixel-level individual typology angle (ITA). They further quantify lesion-to-surrounding-skin contrast using Wasserstein distance. Their findings reveal that global skin lightness exhibits only a weak correlation with segmentation performance, whereas low lesion-skin contrast significantly increases segmentation error, indicating that boundary ambiguity and poor contrast are primary causes of model failure. This work thus offers a novel perspective and quantitative foundation for improving the fairness and robustness of segmentation algorithms across diverse skin tones.

0 citationsRead paper