Benchmarking Off-the-Shelf Multimodal AI Models Against Dermatologists on Patient-Captured Skin Images
研究评估了三种低价至中价的多模态AI模型在患者提交的皮肤图像诊断中的表现,对比了三位认证皮肤科医生的诊断结果,并探讨了模型自信度、额外患者信息和成本对性能的影响。
研究评估了三种低价至中价的多模态AI模型在患者提交的皮肤图像诊断中的表现,对比了三位认证皮肤科医生的诊断结果,并探讨了模型自信度、额外患者信息和成本对性能的影响。
本文提出EffiRAG系统,通过轻量级图构建与查询处理减少成本,同时保持高质量答案生成,适用于需要多文档信息的问题解答。
该论文介绍了Cnuas,一个通过功能仿真解决AI/HPC系统软件开发对昂贵硬件依赖问题的开源平台。
研究通过两个实证研究评估了大型语言模型在需求工程中的应用,覆盖了从需求分类到可追溯性链接识别等五项活动,旨在解决需求信息提取的难题。
This study addresses the vulnerability of resource-constrained devices in healthcare Internet of Things (H-IoT) systems to cyberattacks such as DDoS, man-in-the-middle (MITM), and selective forwarding, which pose serious risks to patient safety. Existing intrusion detection approaches are often hindered by low-quality datasets and computationally intensive models. To overcome these limitations, this work introduces a novel framework that integrates physiological signals with network traffic features and constructs three realistic multi-attack H-IoT datasets using Cooja and ns-3 simulations. A lightweight temporal convolutional network (TCN/Res-TCN) is proposed, augmented with a dynamic thresholding mechanism and optimized monitoring frequency. The model is quantized via TensorFlow Lite and deployed on a Raspberry Pi 4. Experimental results demonstrate real-time attack detection with low latency and power consumption under MQTT/UDP protocols, enabling efficient edge-based security for H-IoT environments.
研究评估了三种低价至中价的多模态AI模型在患者提交的皮肤图像诊断中的表现,对比了三位认证皮肤科医生的诊断结果,并探讨了模型自信度、额外患者信息和成本对性能的影响。
本文提出EffiRAG系统,通过轻量级图构建与查询处理减少成本,同时保持高质量答案生成,适用于需要多文档信息的问题解答。
该论文介绍了Cnuas,一个通过功能仿真解决AI/HPC系统软件开发对昂贵硬件依赖问题的开源平台。
研究通过两个实证研究评估了大型语言模型在需求工程中的应用,覆盖了从需求分类到可追溯性链接识别等五项活动,旨在解决需求信息提取的难题。
This study addresses the vulnerability of resource-constrained devices in healthcare Internet of Things (H-IoT) systems to cyberattacks such as DDoS, man-in-the-middle (MITM), and selective forwarding, which pose serious risks to patient safety. Existing intrusion detection approaches are often hindered by low-quality datasets and computationally intensive models. To overcome these limitations, this work introduces a novel framework that integrates physiological signals with network traffic features and constructs three realistic multi-attack H-IoT datasets using Cooja and ns-3 simulations. A lightweight temporal convolutional network (TCN/Res-TCN) is proposed, augmented with a dynamic thresholding mechanism and optimized monitoring frequency. The model is quantized via TensorFlow Lite and deployed on a Raspberry Pi 4. Experimental results demonstrate real-time attack detection with low latency and power consumption under MQTT/UDP protocols, enabling efficient edge-based security for H-IoT environments.