Robust BiLSTM-Based Multi-Class Intrusion Detection for IoT Networks Using ToN-IoT Dataset

The Internet of Things (IoT) is revolutionizing smart cities, healthcare, and industrial automation, but its devices’ insufficient computing power and weak security make them main targets for advanced cyberattacks, such as DDoS and zero-day exploits. Conventional Intrusion Detection Systems (IDSs) often fail to detect evolving threats, necessitating smarter, deep learning-based solutions. We devise a Bidirectional Long Short-Term Memory (BiLSTM) based IDS trained on the ToN-IoT dataset’s network traffic. Our preprocessing pipeline featuring label encoding, feature hashing, SMOTE for class balancing, and chi-squared feature selection creates a robust, compact feature space. With class weights to address residual imbalance, the model achieves 99.94% accuracy, with high precision, recall, and F1-scores across nine attack types. Its performance remains stable across reduced feature sets, preserving adaptability. Outperforming state-of-the-art methods, our IDS incorporate exceptional detection accuracy with low inference latency, enabling real-time deployment in resource-constrained IoT environments. This framework safely strengthens security for smart cities, critical infrastructure, and edge computing applications.