Selected work

Representative Papers

On the Physical Plausibility and Distribution Alignment for Sim-to-Real RF Positioning

Jul 05, 2026

This work addresses the poor generalization of RF-based localization models to unseen urban streets by pretraining on large-scale synthetic data generated via ray-tracing simulation (using the Sionna platform) over the city of Rome. The study systematically investigates the impact of base station calibration, physical plausibility, dataset scale, and RSSI distribution on simulation-to-reality transfer. It finds that aligning RSSI distributions is more critical than physical fidelity or data volume, leading to an effective distribution normalization strategy. Experiments show that synthetic pretraining consistently improves localization accuracy on known streets, with city-scale unconstrained data yielding the best performance. Crucially, on previously unseen streets, only simulations with aligned RSSI distributions significantly enhance real-world localization accuracy.

0 citationsRead paper

Fusion of Pervasive RF Data with Spatial Images via Vision Transformers for Enhanced Mapping in Smart Cities

Jul 31, 2025

To address the insufficient mapping accuracy of open-source maps—caused by human annotation errors and delayed dynamic updates—this paper proposes a multimodal vision transformer framework that jointly fuses radio frequency (RF) path loss measurements and open-source map imagery. It is the first work to incorporate DINOv2 for urban building mapping, enabling joint modeling of spatial structural priors and RF signal propagation characteristics. The method achieves end-to-end building layout reconstruction solely from aggregated RF path loss data, without requiring expensive remote sensing inputs or manual annotations. Evaluated on a synthetic dataset, it achieves a macro-IoU of 65.3%, substantially outperforming erroneous-map (40.1%), RF-only (37.3%), and non-AI fusion baselines (42.2%). It also surpasses all baselines in Jaccard index, Hausdorff distance, and Chamfer distance. This work establishes a new paradigm for low-cost, high-robustness environmental perception in smart cities.

0 citationsRead paper

On the Limitations of Ray-Tracing for Learning-Based RF Tasks in Urban Environments

Jul 25, 2025

This study addresses the low fidelity of Sionna v1.0.2 in ray-tracing simulations of outdoor cellular links in Rome’s urban environment. We systematically evaluate the impact of geometric modeling, antenna radiation characteristics, and urban residual noise on RF simulation accuracy. A multi-parameter sensitivity analysis framework—incorporating path depth, scattering type, carrier frequency, antenna height, and radiation pattern—is proposed, grounded in real-world measurement data. A lightweight greedy optimization algorithm is designed to jointly calibrate antenna position and orientation. Results show Spearman correlation improvements of 5%–130% across base stations; kNN localization using purely simulated RSSI reduces error by ~33% in real deployments, yet remains 2.1× higher than measured baselines. The core contribution lies in identifying antenna pose as the dominant factor limiting simulation transferability, and in demonstrating that accurate modeling of urban residual noise remains a critical bottleneck for high-fidelity outdoor RF simulation.

0 citationsRead paper

Uncertainty-aware abstention in medical diagnosis based on medical texts

Feb 25, 2025

To address the insufficient reliability of AI models in clinical decision support, this work proposes an uncertainty-aware selective prediction mechanism that enables models to abstain from diagnosis when prediction confidence falls below a task-adaptive threshold. To tackle cross-task uncertainty calibration—particularly challenging across heterogeneous medical tasks—we introduce HUQ-2, a novel uncertainty quantification method integrating Bayesian approximate inference, ensemble-based estimation, and adaptive thresholding. Evaluated on diverse clinical NLP tasks—including in-hospital mortality prediction, multi-label ICD coding recommendation, outpatient triage, and depression/anxiety detection—HUQ-2 is validated on MIMIC-III/IV, proprietary outpatient, and multi-source mental health datasets. It achieves the first unified modeling framework for cross-task, multi-source, heterogeneous clinical text data. Experiments demonstrate a 12.6–18.3% improvement in selective prediction AUC, >94% accuracy on abstained samples, and substantial gains in calibration, robustness, and clinical interpretability—establishing HUQ-2 as the first SOTA uncertainty modeling framework for medical text analysis.

0 citationsRead paper

With Great Backbones Comes Great Adversarial Transferability

Jan 21, 2025

This work investigates the robustness risks of self-supervised learning (SSL)-pretrained vision backbones (ResNet/ViT) under adversarial attacks. Focusing on transferability-induced fragility introduced during fine-tuning, we systematically evaluate 20,000 fine-tuning configurations to uncover latent mechanisms behind robustness degradation. We propose *backbone attack*, a novel black-box method that approximates white-box attack performance using only the frozen backbone—without access to the classifier head. Additionally, we introduce a *surrogate model framework* to quantify how meta-information—specifically architecture, data, and optimization strategy choices—contributes to adversarial transferability. Experiments show that backbone attack significantly outperforms conventional black-box attacks and closely matches white-box performance; surrogate-based attacks achieve high transferability with minimal hyperparameter knowledge; and the relative influence of each fine-tuning dimension is explicitly characterized.

0 citationsRead paper

Recent publications

Latest Papers

On the Physical Plausibility and Distribution Alignment for Sim-to-Real RF Positioning

Jul 05, 2026

This work addresses the poor generalization of RF-based localization models to unseen urban streets by pretraining on large-scale synthetic data generated via ray-tracing simulation (using the Sionna platform) over the city of Rome. The study systematically investigates the impact of base station calibration, physical plausibility, dataset scale, and RSSI distribution on simulation-to-reality transfer. It finds that aligning RSSI distributions is more critical than physical fidelity or data volume, leading to an effective distribution normalization strategy. Experiments show that synthetic pretraining consistently improves localization accuracy on known streets, with city-scale unconstrained data yielding the best performance. Crucially, on previously unseen streets, only simulations with aligned RSSI distributions significantly enhance real-world localization accuracy.

0 citationsRead paper

Fusion of Pervasive RF Data with Spatial Images via Vision Transformers for Enhanced Mapping in Smart Cities

Jul 31, 2025

To address the insufficient mapping accuracy of open-source maps—caused by human annotation errors and delayed dynamic updates—this paper proposes a multimodal vision transformer framework that jointly fuses radio frequency (RF) path loss measurements and open-source map imagery. It is the first work to incorporate DINOv2 for urban building mapping, enabling joint modeling of spatial structural priors and RF signal propagation characteristics. The method achieves end-to-end building layout reconstruction solely from aggregated RF path loss data, without requiring expensive remote sensing inputs or manual annotations. Evaluated on a synthetic dataset, it achieves a macro-IoU of 65.3%, substantially outperforming erroneous-map (40.1%), RF-only (37.3%), and non-AI fusion baselines (42.2%). It also surpasses all baselines in Jaccard index, Hausdorff distance, and Chamfer distance. This work establishes a new paradigm for low-cost, high-robustness environmental perception in smart cities.

0 citationsRead paper

On the Limitations of Ray-Tracing for Learning-Based RF Tasks in Urban Environments

Jul 25, 2025

This study addresses the low fidelity of Sionna v1.0.2 in ray-tracing simulations of outdoor cellular links in Rome’s urban environment. We systematically evaluate the impact of geometric modeling, antenna radiation characteristics, and urban residual noise on RF simulation accuracy. A multi-parameter sensitivity analysis framework—incorporating path depth, scattering type, carrier frequency, antenna height, and radiation pattern—is proposed, grounded in real-world measurement data. A lightweight greedy optimization algorithm is designed to jointly calibrate antenna position and orientation. Results show Spearman correlation improvements of 5%–130% across base stations; kNN localization using purely simulated RSSI reduces error by ~33% in real deployments, yet remains 2.1× higher than measured baselines. The core contribution lies in identifying antenna pose as the dominant factor limiting simulation transferability, and in demonstrating that accurate modeling of urban residual noise remains a critical bottleneck for high-fidelity outdoor RF simulation.

0 citationsRead paper

Uncertainty-aware abstention in medical diagnosis based on medical texts

Feb 25, 2025

To address the insufficient reliability of AI models in clinical decision support, this work proposes an uncertainty-aware selective prediction mechanism that enables models to abstain from diagnosis when prediction confidence falls below a task-adaptive threshold. To tackle cross-task uncertainty calibration—particularly challenging across heterogeneous medical tasks—we introduce HUQ-2, a novel uncertainty quantification method integrating Bayesian approximate inference, ensemble-based estimation, and adaptive thresholding. Evaluated on diverse clinical NLP tasks—including in-hospital mortality prediction, multi-label ICD coding recommendation, outpatient triage, and depression/anxiety detection—HUQ-2 is validated on MIMIC-III/IV, proprietary outpatient, and multi-source mental health datasets. It achieves the first unified modeling framework for cross-task, multi-source, heterogeneous clinical text data. Experiments demonstrate a 12.6–18.3% improvement in selective prediction AUC, >94% accuracy on abstained samples, and substantial gains in calibration, robustness, and clinical interpretability—establishing HUQ-2 as the first SOTA uncertainty modeling framework for medical text analysis.

0 citationsRead paper

With Great Backbones Comes Great Adversarial Transferability

Jan 21, 2025

This work investigates the robustness risks of self-supervised learning (SSL)-pretrained vision backbones (ResNet/ViT) under adversarial attacks. Focusing on transferability-induced fragility introduced during fine-tuning, we systematically evaluate 20,000 fine-tuning configurations to uncover latent mechanisms behind robustness degradation. We propose *backbone attack*, a novel black-box method that approximates white-box attack performance using only the frozen backbone—without access to the classifier head. Additionally, we introduce a *surrogate model framework* to quantify how meta-information—specifically architecture, data, and optimization strategy choices—contributes to adversarial transferability. Experiments show that backbone attack significantly outperforms conventional black-box attacks and closely matches white-box performance; surrogate-based attacks achieve high transferability with minimal hyperparameter knowledge; and the relative influence of each fine-tuning dimension is explicitly characterized.

0 citationsRead paper