Article
SECURE GNNS: DEFENDING GRAPH DATA PRIVACY WITH RANDOMIZED EDGE PERTURBATIONS
The exponential growth of Internet-of-Things (IoT) devices has drastically expanded the threat landscape, yet many existing intrusion detection systems overlook the underlying graph structures within device communication networks. This study proposes a comprehensive framework that integrates traditional machine learning models with a graph-based learning strategy enhanced by randomized edge perturbation, aiming to achieve both high detection performance and adversarial privacy protection. The approach begins with the RT_IOT2022 network-flow dataset, performing exploratory data analysis to uncover feature relationships and address class imbalance. Categorical variables such as attack type, protocol, and service are encoded, and SMOTE oversampling is applied to ensure balanced class distribution. Dimensionality reduction using Principal Component Analysis (PCA) yields a compressed yet information-rich feature space. Four baseline classifiers—Gaussian Naive Bayes (GNB), Linear Discriminant Analysis (LDA), Ridge Classifier, and Decision Tree Classifier (DTC)—are trained and evaluated using standard performance metrics. To exploit graph structural information, we propose a hybrid Graph Neural Network–style Decision Tree Classifier (GNN-DTC), where node features are aggregated from local neighborhoods to construct decision trees that capture relational attack behaviors. Additionally, randomized edge perturbation is applied to mask sensitive graph topology and enhance resilience against adversarial attacks, without significantly compromising accuracy. Experimental results reveal that LDA and Ridge achieve accuracies of 93.4% and 90.8% respectively, while GNB trails with 49.0%. The proposed GNN-DTC model achieves 99.7% accuracy and F1-score, substantially outperforming all baselines. Class-wise evaluations further confirm consistent performance across varied attack types. These findings highlight the effectiveness of combining graph-aware learning with privacy-preserving perturbation for secure and robust IoT intrusion detection
Full Text Attachment
