Explainable and Trust-Oriented Evaluation for Determining Suitable Augmented and Synthetic Data Proportions in IoT...

Artificial-data inclusion can improve IoT intrusion detection, but predictive performance alone cannot establish the trustworthiness of the model. This study presents an explainable and trust-oriented evaluation for determining suitable proportions of augmented and synthetic data in IoT intrusion-detection training. Experiments are conducted on the UM-NIDS and CICIoT datasets using original-data baselines, support vector machine-synthetic minority oversampling technique (SVM-SMOTE) augmentation, and conditional tabular generative adversarial network (CTGAN) synthesis at 33% (low), 50% (medium), and 65% (high) artificial-data proportions. The evaluation considers three complementary trust-oriented dimensions: data quality, predictive reliability, and explanation stability. Data quality is assessed through distributional fidelity, statistical realism, and empirical privacy-risk screening. Predictive reliability is evaluated using conventional classification measures together with imbalance-aware metrics, including Matthews correlation coefficient and Cohen's kappa. Explanation stability is examined using Shapley additive explanations (SHAP), local interpretable model-agnostic explanations (LIME), and integrated gradients (IGs) through Top- $k$ feature overlap, Spearman rank correlation, attribution cosine similarity, and SHAP-specific drift measures relative to the original-data baseline. The findings underscore that suitable proportions of artificial-data inclusion must be determined through dataset- and method-specific evidence rather than assumed universally. This study establishes a systematic basis for selecting such proportions to support reliable, interpretable, and trustworthy IoT intrusion detection.