2026, Vol.12, No.3

Journal of Internet of Things and Convergence 2024 KCI Impact Factor : 0.69

pISSN : 2466-0078
https://journal.kci.go.kr/kiots

Research is actively underway to establish automated environments through the application of artificial intelligence across various fields. However, building high-specification computing environments for applying machine learning incurs significant costs. Consequently, the need for research on efficiently applying AI in low-specification environments has emerged. Machine learning on high-dimensional data in low-specification computing environments suffers severe performance degradation due to memory constraints and computational costs. This study proposes a dimension reduction technique utilizing Principal Component Analysis (PCA) to address these issues. The proposed method adds a preprocessing step using PCA before performing machine learning on high-dimensional data, enabling machine learning to be performed using the refined data. This simultaneously improves the performance and efficiency of machine learning models even in environments with limited computing resources. Experiments confirmed that PCA-based preprocessing effectively reduces the dimensions of multidimensional data while preserving key variance information under identical conditions. Consequently, average training time decreased by 96.6%, memory usage was reduced by 99.7%, and model prediction accuracy remained at a comparable level. This research demonstrates that efficient machine learning implementation is possible even in environments with limited hardware resources, such as embedded systems or edge computing. This is expected to provide a practical solution for researchers and developers with limited access to high-performance computing infrastructure.

With the increasing security threats from the proliferation of IP cameras, the importance of machine learning-based anomaly detection is growing. However, existing studies are often biased toward single datasets, limiting their generalization performance in heterogeneous network environments. This study proposes a robust Stacking ensemble-based intrusion detection model by cross-utilizing the host-based N-BaIoT and network-based Kitsune datasets. Specifically, RF, NN, and AdaBoost were configured as base learners, and hyperparameter optimization was performed in the Orange3 environment. Experimental results show that the proposed model achieves an AUC greater than 0.999 under the intra-validation scenario. In inter-validation scenarios, performance degradation is observed due to differences in data characteristics. Information Gain analysis reveals that the asymmetric anomaly signals between the strong attack patterns in N-BaIoT and the relatively subtle anomalies in Kitsune are the primary cause of this performance degradation. Furthermore, the proposed stacking model demonstrates more robust performance than single machine learning models, suggesting its applicability to real-world security solutions where heterogeneous data sources coexist.

Recent advances in artificial intelligence technology have driven the development of automated systems across various fields. However, acquiring large-scale, high-quality data required for training high-performance machine learning models demands significant costs and time. Particularly in specific domains or resource-constrained environments, the amount of collectable data is limited, leading to serious issues such as model overfitting and reduced generalization performance. This study proposes a data management and learning policy that systematically applies data augmentation techniques to address this data scarcity problem. The proposed method enhances the diversity of the training dataset by applying various augmentation techniques, such as geometric transformations and noise insertion, based on the limited original data. This approach overcomes the ‘curse of dimensionality’ and improves the model's prediction accuracy. This study demonstrates that efficient machine learning implementation is achievable even in environments with limited data resources through data augmentation policies. This is expected to provide a practical solution for researchers with restricted access to high-performance computing infrastructure or large-scale datasets.