SPADE: SPaT Attack Detection from the Connected Vehicle's Perspective
为解决车联网中SPaT消息的完整性受到威胁的问题，本文通过创建一个多模态仿真数据集SPADE来支持深度学习入侵检测系统的研究。
0 citationsRead paper
Selected work
Representative Papers
Energy and Performance Benchmarking of Deep Learning Models for Breast Cancer Detection
This study addresses the high computational complexity of deep learning models in breast cancer detection, which incurs substantial energy consumption and carbon emissions, necessitating a balance between performance and environmental sustainability. For the first time in this domain, we jointly evaluate the accuracy and CO₂ emissions of seven mainstream architectures—including CNNs, Transformers, and hybrid models—on the Breast Ultrasound and BreakHis 400X datasets. Our results demonstrate that DeiT achieves the best trade-off between efficiency and accuracy on ultrasound data, while ViT and Swin perform optimally on BreakHis. The absence of a universally superior architecture across both modalities underscores the importance of multidimensional model selection that accounts for both data characteristics and environmental impact.
0 citationsRead paper
Towards Sustainable Artificial Intelligence: A Comprehensive Review and Comparative Analysis of Deep Learning Models' Carbon Footprint
This study addresses the substantial carbon emissions associated with the high energy consumption of deep learning models, calling for systematic evaluation and mitigation of their environmental impact. Combining a systematic literature review with empirical analysis, it presents the first multi-label classification experiments on six deep learning models conducted on CPU platforms, quantifying their full lifecycle carbon footprints and benchmarking widely used carbon accounting tools. The findings reveal that the training phase dominates emissions and that model complexity exhibits a nonlinear relationship with accuracy gains—increased architectural sophistication does not consistently yield proportional performance improvements. These results provide empirical guidance for model selection that balances predictive accuracy with environmental sustainability, thereby advancing the paradigm of green AI design.
0 citationsRead paper
Do Transformers Actually Help Intrusion Detection? A Temporal Sequence Evaluation on CIC-IDS2017
This study addresses the pervasive issue of performance overestimation in network intrusion detection research due to data leakage and unrealistic sequential inputs during temporal model evaluation. The authors reformulate the CIC-IDS2017 dataset as a realistic network session sequencing task and introduce a leakage-free data partitioning scheme alongside multiple padding strategies. They systematically evaluate the temporal modeling capabilities of nine architectures—including Transformer, LSTM, GRU, 1D-CNN, and Random Forest—revealing for the first time that data partitioning and padding choices exert a far greater impact on performance than model architecture itself. Specifically, random splitting with repetition padding substantially inflates robustness estimates; under authentic sequential windows, the Transformer achieves the highest macro-F1 (0.89), yet drops by 0.24 with zero-padding plus masking, while Random Forest demonstrates superior robustness in leakage-free evaluation, and the Transformer’s false positive rate surges 67-fold.
0 citationsRead paper
Is this Build Failure Related to my Patch? An Empirical Study of Unrelated Build Failures in Continuous Integration
This study addresses the challenge developers face in continuous integration (CI) environments when determining whether a build failure stems from their own code changes, often leading to costly and unnecessary debugging efforts. The work presents the first systematic identification and categorization of “irrelevant build failures”—failures unrelated to a developer’s recent commit. Through an empirical analysis of 77,354 failed builds across seven Apache projects, combined with documentation mining and a rich set of 33 features—including CI latency, error recurrence rate, and comment count—the authors propose a prediction approach based on Positive and Unlabeled (PU) learning. Evaluated across multiple projects, the model achieves precision of 0.70–0.88, F1-scores of 0.44–0.91, and AUC values of 0.63–0.97, significantly improving the efficiency of identifying irrelevant failures and alleviating developers’ diagnostic burden.
0 citationsRead paper
Recent publications
Latest Papers
SPADE: SPaT Attack Detection from the Connected Vehicle's Perspective
为解决车联网中SPaT消息的完整性受到威胁的问题，本文通过创建一个多模态仿真数据集SPADE来支持深度学习入侵检测系统的研究。
0 citationsRead paper
Energy and Performance Benchmarking of Deep Learning Models for Breast Cancer Detection
This study addresses the high computational complexity of deep learning models in breast cancer detection, which incurs substantial energy consumption and carbon emissions, necessitating a balance between performance and environmental sustainability. For the first time in this domain, we jointly evaluate the accuracy and CO₂ emissions of seven mainstream architectures—including CNNs, Transformers, and hybrid models—on the Breast Ultrasound and BreakHis 400X datasets. Our results demonstrate that DeiT achieves the best trade-off between efficiency and accuracy on ultrasound data, while ViT and Swin perform optimally on BreakHis. The absence of a universally superior architecture across both modalities underscores the importance of multidimensional model selection that accounts for both data characteristics and environmental impact.
0 citationsRead paper
Towards Sustainable Artificial Intelligence: A Comprehensive Review and Comparative Analysis of Deep Learning Models' Carbon Footprint
This study addresses the substantial carbon emissions associated with the high energy consumption of deep learning models, calling for systematic evaluation and mitigation of their environmental impact. Combining a systematic literature review with empirical analysis, it presents the first multi-label classification experiments on six deep learning models conducted on CPU platforms, quantifying their full lifecycle carbon footprints and benchmarking widely used carbon accounting tools. The findings reveal that the training phase dominates emissions and that model complexity exhibits a nonlinear relationship with accuracy gains—increased architectural sophistication does not consistently yield proportional performance improvements. These results provide empirical guidance for model selection that balances predictive accuracy with environmental sustainability, thereby advancing the paradigm of green AI design.
0 citationsRead paper
Do Transformers Actually Help Intrusion Detection? A Temporal Sequence Evaluation on CIC-IDS2017
This study addresses the pervasive issue of performance overestimation in network intrusion detection research due to data leakage and unrealistic sequential inputs during temporal model evaluation. The authors reformulate the CIC-IDS2017 dataset as a realistic network session sequencing task and introduce a leakage-free data partitioning scheme alongside multiple padding strategies. They systematically evaluate the temporal modeling capabilities of nine architectures—including Transformer, LSTM, GRU, 1D-CNN, and Random Forest—revealing for the first time that data partitioning and padding choices exert a far greater impact on performance than model architecture itself. Specifically, random splitting with repetition padding substantially inflates robustness estimates; under authentic sequential windows, the Transformer achieves the highest macro-F1 (0.89), yet drops by 0.24 with zero-padding plus masking, while Random Forest demonstrates superior robustness in leakage-free evaluation, and the Transformer’s false positive rate surges 67-fold.
0 citationsRead paper
Is this Build Failure Related to my Patch? An Empirical Study of Unrelated Build Failures in Continuous Integration
This study addresses the challenge developers face in continuous integration (CI) environments when determining whether a build failure stems from their own code changes, often leading to costly and unnecessary debugging efforts. The work presents the first systematic identification and categorization of “irrelevant build failures”—failures unrelated to a developer’s recent commit. Through an empirical analysis of 77,354 failed builds across seven Apache projects, combined with documentation mining and a rich set of 33 features—including CI latency, error recurrence rate, and comment count—the authors propose a prediction approach based on Positive and Unlabeled (PU) learning. Evaluated across multiple projects, the model achieves precision of 0.70–0.88, F1-scores of 0.44–0.91, and AUC values of 0.63–0.97, significantly improving the efficiency of identifying irrelevant failures and alleviating developers’ diagnostic burden.
0 citationsRead paper