Selected work

Representative Papers

AI-Driven Radiology Report Generation for Traumatic Brain Injuries.

Jan 30, 2025Journal of imaging informatics in medicine

To address diagnostic delays caused by delayed interpretation of cranial trauma imaging in emergency settings, this study proposes an end-to-end AI system integrating AC-BiFPN and Transformer architectures for multi-scale feature extraction from CT/MRI scans and automatic generation of natural-language radiology reports. The framework uniquely co-optimizes lesion detection accuracy and report semantic coherence in the cranial trauma domain: AC-BiFPN enhances multi-scale lesion localization, while the Transformer captures long-range semantic dependencies to produce structured, clinically interpretable reports. Evaluated on the RSNA Intracranial Hemorrhage dataset, the model achieves significantly higher diagnostic accuracy and report quality compared to conventional CNN-based approaches. This work advances emergency department efficiency and provides an interpretable, deployable solution for clinical decision support and medical education.

3 citationsRead paper

Dynamic Mask-Based Backdoor Attack Against Vision AI Models: A Case Study on Mushroom Detection

Jan 26, 2026

This work proposes a novel backdoor attack method based on dynamic masks to address the vulnerability of vision AI models to data poisoning. By leveraging dynamic masks generated by the Segment Anything Model (SAM) as adaptive triggers, the approach embeds these triggers into the training data of an object detection model (YOLOv7), overcoming the limitations of conventional static triggers. Experimental results demonstrate that the proposed method achieves over 90% attack success rate on poisoned samples while preserving high detection accuracy on clean samples, thereby significantly enhancing both the stealthiness and effectiveness of the backdoor attack.

0 citationsRead paper

Enhancing Early Alzheimer Disease Detection through Big Data and Ensemble Few-Shot Learning

Oct 22, 2025

Addressing critical challenges in early Alzheimer’s disease (AD) detection—including severe scarcity of labeled data, complex neuropathology, and stringent medical data privacy constraints—this paper proposes a prototype-based few-shot ensemble deep learning framework. Methodologically, it integrates multi-source pretrained CNNs to extract multi-scale features from medical neuroimaging; further, it introduces a novel joint optimization mechanism combining class-aware loss and entropy regularization to enhance discriminability and generalizability under few-shot conditions. Evaluated on the Kaggle Alzheimer and ADNI public benchmarks, the model achieves classification accuracies of 99.72% and 99.86%, respectively—substantially outperforming state-of-the-art methods. This work establishes a scalable, privacy-preserving technical paradigm for accurate, low-label-cost early AD screening.

0 citationsRead paper

FedMicro-IDA: A Federated Learning and Microservices-based Framework for IoT Data Analytics

Oct 22, 2025

To address the challenge of efficiently analyzing multi-source heterogeneous IoT data under stringent constraints of privacy preservation, low latency, and high reliability, this paper proposes the first edge intelligence analytics framework integrating federated learning with a microservice architecture. The framework enables dynamic orchestration and reuse of intelligent microservices, unifying collaborative model training and elastic system scalability. It employs a Docker/Kubernetes-based microservice deployment infrastructure augmented by edge computing to minimize communication latency and bandwidth consumption. Experimental evaluation on the MaleVis dataset demonstrates that the proposed method achieves a malware detection accuracy of 99.24%, significantly outperforming existing approaches. Crucially, it ensures data locality and model privacy security throughout the learning process. The results validate the framework’s comprehensive advantages in analytical performance, scalability, and privacy protection.

0 citationsRead paper

PrompTrend: Continuous Community-Driven Vulnerability Discovery and Assessment for Large Language Models

Jul 25, 2025

Static benchmarking inadequately captures LLM security vulnerabilities exposed in real-world online community practices. To address this, we propose a community-driven dynamic monitoring paradigm, focusing on psychological attacks as the primary threat vector, revealing that capability advancement and safety improvement are misaligned. Methodologically, we design a cross-platform data collection system, a multidimensional scoring framework, and a scalable monitoring architecture—integrating horizontal comparative analysis with fine-grained vulnerability classification. Over five months, we conduct an empirical study across nine commercial LLMs. Our approach identifies 198 novel vulnerabilities with 78% classification accuracy; psychological attacks exhibit significantly higher detection rates than traditional exploit-based techniques yet demonstrate low cross-model transferability—highlighting their stealthiness and model specificity. This work pioneers systematic, continuous discovery and quantitative evaluation of community-emergent LLM vulnerabilities.

0 citationsRead paper

Recent publications

Latest Papers

Dynamic Mask-Based Backdoor Attack Against Vision AI Models: A Case Study on Mushroom Detection

Jan 26, 2026

This work proposes a novel backdoor attack method based on dynamic masks to address the vulnerability of vision AI models to data poisoning. By leveraging dynamic masks generated by the Segment Anything Model (SAM) as adaptive triggers, the approach embeds these triggers into the training data of an object detection model (YOLOv7), overcoming the limitations of conventional static triggers. Experimental results demonstrate that the proposed method achieves over 90% attack success rate on poisoned samples while preserving high detection accuracy on clean samples, thereby significantly enhancing both the stealthiness and effectiveness of the backdoor attack.

0 citationsRead paper

Enhancing Early Alzheimer Disease Detection through Big Data and Ensemble Few-Shot Learning

Oct 22, 2025

Addressing critical challenges in early Alzheimer’s disease (AD) detection—including severe scarcity of labeled data, complex neuropathology, and stringent medical data privacy constraints—this paper proposes a prototype-based few-shot ensemble deep learning framework. Methodologically, it integrates multi-source pretrained CNNs to extract multi-scale features from medical neuroimaging; further, it introduces a novel joint optimization mechanism combining class-aware loss and entropy regularization to enhance discriminability and generalizability under few-shot conditions. Evaluated on the Kaggle Alzheimer and ADNI public benchmarks, the model achieves classification accuracies of 99.72% and 99.86%, respectively—substantially outperforming state-of-the-art methods. This work establishes a scalable, privacy-preserving technical paradigm for accurate, low-label-cost early AD screening.

0 citationsRead paper

FedMicro-IDA: A Federated Learning and Microservices-based Framework for IoT Data Analytics

Oct 22, 2025

To address the challenge of efficiently analyzing multi-source heterogeneous IoT data under stringent constraints of privacy preservation, low latency, and high reliability, this paper proposes the first edge intelligence analytics framework integrating federated learning with a microservice architecture. The framework enables dynamic orchestration and reuse of intelligent microservices, unifying collaborative model training and elastic system scalability. It employs a Docker/Kubernetes-based microservice deployment infrastructure augmented by edge computing to minimize communication latency and bandwidth consumption. Experimental evaluation on the MaleVis dataset demonstrates that the proposed method achieves a malware detection accuracy of 99.24%, significantly outperforming existing approaches. Crucially, it ensures data locality and model privacy security throughout the learning process. The results validate the framework’s comprehensive advantages in analytical performance, scalability, and privacy protection.

0 citationsRead paper

PrompTrend: Continuous Community-Driven Vulnerability Discovery and Assessment for Large Language Models

Jul 25, 2025

Static benchmarking inadequately captures LLM security vulnerabilities exposed in real-world online community practices. To address this, we propose a community-driven dynamic monitoring paradigm, focusing on psychological attacks as the primary threat vector, revealing that capability advancement and safety improvement are misaligned. Methodologically, we design a cross-platform data collection system, a multidimensional scoring framework, and a scalable monitoring architecture—integrating horizontal comparative analysis with fine-grained vulnerability classification. Over five months, we conduct an empirical study across nine commercial LLMs. Our approach identifies 198 novel vulnerabilities with 78% classification accuracy; psychological attacks exhibit significantly higher detection rates than traditional exploit-based techniques yet demonstrate low cross-model transferability—highlighting their stealthiness and model specificity. This work pioneers systematic, continuous discovery and quantitative evaluation of community-emergent LLM vulnerabilities.

0 citationsRead paper

Bridging AI and Software Security: A Comparative Vulnerability Assessment of LLM Agent Deployment Paradigms

Jul 08, 2025

Current LLM agent security research treats AI-specific vulnerabilities and traditional software flaws in isolation, lacking a unified cross-domain evaluation framework. Method: This work presents the first systematic comparison of two dominant deployment paradigms—Function Calling (FC) and Model Context Protocol (MCP)—and introduces a unified threat taxonomy integrating AI reasoning vulnerabilities with classical software security concepts. We empirically evaluate the robustness of seven LLMs against prompt injection, JSON injection, DoS, and multi-step chained attacks across 3,250 adversarial scenarios. Contribution/Results: FC exhibits higher overall attack success (73.5% vs. 62.59%), but risks concentrate at the system layer; MCP exacerbates LLM centralization and exposure. Chained attacks achieve 91–96% success rates, and advanced reasoning models show heightened exploitability. Our findings reveal how architectural choices fundamentally reshape the threat landscape, establishing novel conceptual insights and practical benchmarks for secure LLM agent design.

0 citationsRead paper