همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Lightweight Hybrid Attention Recurrent Neural Network (HARN) for IDS in IoT Network Layer

نویسندگان :

Mohammad Hossein Esfahani¹ Ali Sadr²

1- Iran University of Science and Technology (IUST) Tehran, Iran 2- Iran University of Science and Technology (IUST) Tehran, Iran

کلمات کلیدی :

TON-IoT،IoT،Intrusion Detection System،Deep Learning،Zero-Day Attack،HARN

چکیده :

The rapid rise of the Internet of Things (IoT) has brought convenience and intelligence to many real-world applications, but it has also opened the door to increasingly sophisticated cyberattacks. Among them, zero-day attacks remain especially difficult to detect because they do not follow recognizable patterns and often bypass traditional intrusion detection systems (IDS). To address this issue, we introduce a lightweight Hybrid Attention Recurrent Neural Network (HARN) that combines LSTM, GRU, and attention layers to better capture the most informative patterns in network traffic. A key strength of the proposed approach is its carefully designed preprocessing stage: statistical feature extraction, Yeo–Johnson transformation, SMOTE rebalancing, and RFE-based feature selection work together to form a compact and clean feature set that noticeably improves the model’s learning quality and metric outcomes. By setting the recurrent time-step to 1, the model avoids the accumulation of long-term memory and becomes more responsive to previously unseen attack behaviors. Experiments on the TON-IoT dataset show that the proposed IDS achieves 99.67% accuracy, 99.55% precision, 99.75% recall, a 99.71% F1-score, and an AUC of 0.9984, outperforming prior methods on all metrics. These results demonstrate that the combination of the hybrid architecture and the strengthened preprocessing pipeline provides a reliable and practical solution for improving security in modern IoT networks.