U-PEN Mamba: Progressive Expansion with Selective State-Space Modeling for Efficient Retinal Vessel Segmentation
本文提出U-PEN Mamba，通过渐进非线性特征扩展与选择性状态空间建模解决视网膜血管分割中细小血管、低对比度和前景背景不平衡的问题。
0 citationsRead paper
Selected work
Representative Papers
Enhancing Smart Farming Through Federated Learning: A Secure, Scalable, and Efficient Approach for AI-Driven Agriculture
Reconciling farm data privacy preservation with cross-domain collaborative modeling remains challenging in smart agriculture. Method: This paper proposes a lightweight federated learning framework tailored for heterogeneous farmland environments in Minnesota. It integrates transfer learning with a modified Federated Averaging (FedAvg) algorithm, enabling local model training on edge devices and uploading only encrypted gradient updates—never raw image data—to a central server. Dynamic client selection, sparse communication compression, and an edge-optimized lightweight CNN backbone further reduce bandwidth and computational overhead. Contribution/Results: Evaluated on multi-farm heterogeneous datasets, the framework achieves 92.3% accuracy in crop disease classification—only 1.2 percentage points below a centralized baseline—while reducing communication costs by 64%. To our knowledge, this is the first deployable federated crop disease detection solution that simultaneously ensures strong privacy guarantees, low resource dependency, and high generalization performance.
0 citationsRead paper
Fine-Tuning Federated Learning-Based Intrusion Detection Systems for Transportation IoT
To address the challenges of resource constraints, privacy sensitivity, and real-time responsiveness in intrusion detection for vehicular edge devices within traffic-oriented IoT, this paper proposes a lightweight server-edge collaborative federated learning framework. Methodologically, it introduces a novel “pre-training–edge fine-tuning” hybrid paradigm, integrating model pruning and quantization, lightweight neural network architecture design, and a distributed secure evaluation protocol—ensuring data privacy while enabling efficient distributed deployment. Experimental results demonstrate a 42% reduction in memory footprint, a 75% decrease in training time, an intrusion detection system (IDS) accuracy of 99.2%, and less than 1.3% performance degradation at a scale of one thousand nodes. These outcomes significantly enhance scalability and practical applicability in resource-constrained vehicular edge environments.
0 citationsRead paper
Recent publications
Latest Papers
U-PEN Mamba: Progressive Expansion with Selective State-Space Modeling for Efficient Retinal Vessel Segmentation
本文提出U-PEN Mamba，通过渐进非线性特征扩展与选择性状态空间建模解决视网膜血管分割中细小血管、低对比度和前景背景不平衡的问题。
0 citationsRead paper
Enhancing Smart Farming Through Federated Learning: A Secure, Scalable, and Efficient Approach for AI-Driven Agriculture
Reconciling farm data privacy preservation with cross-domain collaborative modeling remains challenging in smart agriculture. Method: This paper proposes a lightweight federated learning framework tailored for heterogeneous farmland environments in Minnesota. It integrates transfer learning with a modified Federated Averaging (FedAvg) algorithm, enabling local model training on edge devices and uploading only encrypted gradient updates—never raw image data—to a central server. Dynamic client selection, sparse communication compression, and an edge-optimized lightweight CNN backbone further reduce bandwidth and computational overhead. Contribution/Results: Evaluated on multi-farm heterogeneous datasets, the framework achieves 92.3% accuracy in crop disease classification—only 1.2 percentage points below a centralized baseline—while reducing communication costs by 64%. To our knowledge, this is the first deployable federated crop disease detection solution that simultaneously ensures strong privacy guarantees, low resource dependency, and high generalization performance.
0 citationsRead paper
Fine-Tuning Federated Learning-Based Intrusion Detection Systems for Transportation IoT
To address the challenges of resource constraints, privacy sensitivity, and real-time responsiveness in intrusion detection for vehicular edge devices within traffic-oriented IoT, this paper proposes a lightweight server-edge collaborative federated learning framework. Methodologically, it introduces a novel “pre-training–edge fine-tuning” hybrid paradigm, integrating model pruning and quantization, lightweight neural network architecture design, and a distributed secure evaluation protocol—ensuring data privacy while enabling efficient distributed deployment. Experimental results demonstrate a 42% reduction in memory footprint, a 75% decrease in training time, an intrusion detection system (IDS) accuracy of 99.2%, and less than 1.3% performance degradation at a scale of one thousand nodes. These outcomes significantly enhance scalability and practical applicability in resource-constrained vehicular edge environments.
0 citationsRead paper