Preprocessing-Enhanced Deep Learning for IoT Intrusion Detection

Abstract

The rapid proliferation of Internet of Things (IoT) devices introduced significant security challenges. Prior machine and deep learning approaches for intrusion detection often exhibited suboptimal performance due to insufficient preprocessing. We proposed a comprehensive preprocessing framework that integrated advanced feature selection, conservative data augmentation, and an Enhanced Model Selection Algorithm. Specifically, we combined correlation-based filtering with mutual information to reduce redundancy while preserving discriminative indicators, applied SMOTE and small-magnitude noise injection to mitigate class imbalance without distorting protocol-consistent patterns, and introduced a model selection algorithm to balance accuracy, recall, and efficiency across ANN, 1D-CNN, and LSTM models. Experiments on the IoTID20 dataset showed that the preprocessing-enhanced framework achieved accuracy above 98% and improved recall compared with baselines, with the CNN delivering the best trade-off between robustness and efficiency. These results demonstrated that systematic preprocessing and model selection materially strengthened IDS performance and highlighted the practicality of deploying compact deep learning models in IoT environments.