Selected work

Representative Papers

SmartGraphical: A Human-in-the-Loop Framework for Detecting Smart Contract Logical Vulnerabilities via Pattern-Driven Static Analysis and Visual Abstraction

Mar 09, 2026

Existing approaches to smart contract vulnerability detection are largely confined to syntactic analysis and struggle to identify deep logical flaws stemming from business logic defects. This work proposes a human-in-the-loop framework that integrates pattern-driven static analysis with visual abstraction, enabling developers to interactively explore logical attack surfaces through functional control flow graphs. By systematically incorporating visualization and expert judgment into the detection process, the approach overcomes the contextual understanding limitations inherent in purely automated tools. Evaluated on a large-scale dataset of real-world contracts and a user study involving 100 developers, the framework not only successfully reproduces high-severity vulnerabilities such as the SYFI rebase failure but also uncovers multiple logic flaws missed by mainstream detection tools, significantly improving both explainability and recall.

0 citationsRead paper

PokeFusion Attention: Enhancing Reference-Free Style-Conditioned Generation

Feb 03, 2026

This work addresses the challenge of simultaneously preserving structural stability and fine-grained style consistency in character generation using text-to-image diffusion models under no-reference conditions. To this end, we propose a lightweight cross-attention mechanism within the diffusion decoder that decouples and fuses textual semantics with learnable style embeddings. By fine-tuning only the decoder’s cross-attention layers and a compact style projection module, our approach achieves parameter-efficient, plug-and-play, and backbone-agnostic style control. Evaluated on the Pokémon character generation benchmark, the method significantly improves style fidelity, semantic alignment, and shape consistency while maintaining low computational overhead and inference simplicity.

0 citationsRead paper

AIS-CycleGen: A CycleGAN-Based Framework for High-Fidelity Synthetic AIS Data Generation and Augmentation

Jan 04, 2026arXiv.org

This study addresses the limitations imposed by domain shift, data sparsity, and class imbalance in Automatic Identification System (AIS) data, which hinder the performance of maritime prediction models. To overcome these challenges, the work proposes an unsupervised domain translation framework based on CycleGAN for AIS trajectory augmentation, eliminating the need for paired training data. The approach integrates a 1D convolutional generator with adaptive noise injection to synthesize high-fidelity trajectories while preserving their spatiotemporal structure. Experimental results demonstrate that the generated data significantly enhance both diversity and realism, leading to superior performance across multiple regression baselines, achieving a PSNR of 30.5 and an FID of 38.9—outperforming existing GAN-based augmentation methods.

0 citationsRead paper

Adaptive Optimizable Gaussian Process Regression Linear Least Squares Regression Filtering Method for SEM Images

Oct 09, 2025

Accurate estimation of signal-to-noise ratio (SNR) and noise variance (NV) in scanning electron microscopy (SEM) images is hindered by severe noise interference, limiting denoising performance. To address this, we propose an NV-guided adaptive Wiener filtering enhancement method. Our approach introduces the AO-GPRLLSR framework—integrating linear least-squares regression (LLSR) with optimizable Gaussian process regression (GPR)—to achieve precise, adaptive NV estimation directly from noisy SEM data. Leveraging the estimated NV, Wiener filter parameters are dynamically configured to maximize denoising robustness. Experimental results demonstrate that our method significantly reduces mean squared error compared to state-of-the-art techniques, achieves superior accuracy in both SNR and NV estimation, and markedly improves SEM image clarity and the reliability of subsequent quantitative analysis.

0 citationsRead paper

GastroViT: A Vision Transformer Based Ensemble Learning Approach for Gastrointestinal Disease Classification with Grad CAM & SHAP Visualization

Sep 30, 2025

This study addresses the challenge of fine-grained classification of gastrointestinal endoscopic images by proposing GastroViT, a lightweight and interpretable ensemble model. Methodologically, it integrates two pretrained vision transformer backbones—MobileViT_XS and MobileViT_V2_200—and enhances feature discriminability via attention mechanisms. Dual-path interpretability is achieved through synergistic Grad-CAM (for lesion localization) and SHAP (for quantitative feature attribution). On 23-class and 16-class gastrointestinal disease classification tasks, GastroViT achieves 91.98% (F1 = 64%) and 92.70% (F1 = 87%) accuracy, respectively, with only 20 million parameters, no data augmentation, and robustness to class imbalance. Its core contribution lies in being the first to jointly leverage Vision Transformer ensembling and multimodal attribution-based visualization—thereby unifying high accuracy, low computational complexity, and clinical interpretability—establishing a novel paradigm for AI-assisted early screening in gastroenterology.

0 citationsRead paper

Recent publications

Latest Papers

SmartGraphical: A Human-in-the-Loop Framework for Detecting Smart Contract Logical Vulnerabilities via Pattern-Driven Static Analysis and Visual Abstraction

Mar 09, 2026

Existing approaches to smart contract vulnerability detection are largely confined to syntactic analysis and struggle to identify deep logical flaws stemming from business logic defects. This work proposes a human-in-the-loop framework that integrates pattern-driven static analysis with visual abstraction, enabling developers to interactively explore logical attack surfaces through functional control flow graphs. By systematically incorporating visualization and expert judgment into the detection process, the approach overcomes the contextual understanding limitations inherent in purely automated tools. Evaluated on a large-scale dataset of real-world contracts and a user study involving 100 developers, the framework not only successfully reproduces high-severity vulnerabilities such as the SYFI rebase failure but also uncovers multiple logic flaws missed by mainstream detection tools, significantly improving both explainability and recall.

0 citationsRead paper

PokeFusion Attention: Enhancing Reference-Free Style-Conditioned Generation

Feb 03, 2026

This work addresses the challenge of simultaneously preserving structural stability and fine-grained style consistency in character generation using text-to-image diffusion models under no-reference conditions. To this end, we propose a lightweight cross-attention mechanism within the diffusion decoder that decouples and fuses textual semantics with learnable style embeddings. By fine-tuning only the decoder’s cross-attention layers and a compact style projection module, our approach achieves parameter-efficient, plug-and-play, and backbone-agnostic style control. Evaluated on the Pokémon character generation benchmark, the method significantly improves style fidelity, semantic alignment, and shape consistency while maintaining low computational overhead and inference simplicity.

0 citationsRead paper

AIS-CycleGen: A CycleGAN-Based Framework for High-Fidelity Synthetic AIS Data Generation and Augmentation

Jan 04, 2026arXiv.org

This study addresses the limitations imposed by domain shift, data sparsity, and class imbalance in Automatic Identification System (AIS) data, which hinder the performance of maritime prediction models. To overcome these challenges, the work proposes an unsupervised domain translation framework based on CycleGAN for AIS trajectory augmentation, eliminating the need for paired training data. The approach integrates a 1D convolutional generator with adaptive noise injection to synthesize high-fidelity trajectories while preserving their spatiotemporal structure. Experimental results demonstrate that the generated data significantly enhance both diversity and realism, leading to superior performance across multiple regression baselines, achieving a PSNR of 30.5 and an FID of 38.9—outperforming existing GAN-based augmentation methods.

0 citationsRead paper

Adaptive Optimizable Gaussian Process Regression Linear Least Squares Regression Filtering Method for SEM Images

Oct 09, 2025

Accurate estimation of signal-to-noise ratio (SNR) and noise variance (NV) in scanning electron microscopy (SEM) images is hindered by severe noise interference, limiting denoising performance. To address this, we propose an NV-guided adaptive Wiener filtering enhancement method. Our approach introduces the AO-GPRLLSR framework—integrating linear least-squares regression (LLSR) with optimizable Gaussian process regression (GPR)—to achieve precise, adaptive NV estimation directly from noisy SEM data. Leveraging the estimated NV, Wiener filter parameters are dynamically configured to maximize denoising robustness. Experimental results demonstrate that our method significantly reduces mean squared error compared to state-of-the-art techniques, achieves superior accuracy in both SNR and NV estimation, and markedly improves SEM image clarity and the reliability of subsequent quantitative analysis.

0 citationsRead paper

GastroViT: A Vision Transformer Based Ensemble Learning Approach for Gastrointestinal Disease Classification with Grad CAM & SHAP Visualization

Sep 30, 2025

This study addresses the challenge of fine-grained classification of gastrointestinal endoscopic images by proposing GastroViT, a lightweight and interpretable ensemble model. Methodologically, it integrates two pretrained vision transformer backbones—MobileViT_XS and MobileViT_V2_200—and enhances feature discriminability via attention mechanisms. Dual-path interpretability is achieved through synergistic Grad-CAM (for lesion localization) and SHAP (for quantitative feature attribution). On 23-class and 16-class gastrointestinal disease classification tasks, GastroViT achieves 91.98% (F1 = 64%) and 92.70% (F1 = 87%) accuracy, respectively, with only 20 million parameters, no data augmentation, and robustness to class imbalance. Its core contribution lies in being the first to jointly leverage Vision Transformer ensembling and multimodal attribution-based visualization—thereby unifying high accuracy, low computational complexity, and clinical interpretability—establishing a novel paradigm for AI-assisted early screening in gastroenterology.

0 citationsRead paper