IoT Device Intrusion Detection Method Based on Improved CNN and Data Balancing Algorithm

Abstract A deep learning model that combines One-dimensional Densely Connected Convolutional Network (DenseNet1D) and Bidirectional-Simple-Recurrent-Unit (Bi-SRU) is proposed to address the spatiotemporal feature extraction and data imbalance in intrusion detection of IoT devices. This model is used to capture the spatiotemporal features of traffic data. Meanwhile, a data balancing algorithm combining Wasserstein Generative Adversarial Network (WGAN) and Adaptive Synthetic Minority oversampling (ADASYN) is designed to solve the sample imbalance. The experimental results showed that this method achieved an accuracy of 98.06%, a recall of 96.84%, and an F1-score of 97.21% on the Bot-IoT dataset, significantly improving the detection performance of minority class samples and verifying its generalization ability to scarce attack patterns. In addition, the average training time was only 3,138 seconds, and the overall detection accuracy was 99.24%, indicating that the model had strong robustness to changes in data distribution. The research proves that the combination of DenseNet1D and Bi-SRU can effectively solve the defect that a single model cannot take into account the spatiotemporal characteristics, which is conducive to extracting complex spatiotemporal characteristics. In addition, the WGAN-ADASYN algorithm significantly improves data imbalance and provides a practical technical solution for real-time security protection of IoT devices in resource constrained scenarios.