Hybrid deep learning-based IoT intrusion detection : A comparative study of CNN, GRU, LSTM, and hybrid architectures

Cyber-physical systems, particularly Internet of Things devices, pose significant cybersecurity challenges due to their vast volume, speed, and complexity of network traffic and attack vectors. This research presents an innovative hybrid deep learning technique to improve intrusion detection by taking full advantage of spatial and temporal characteristics from IoT network traffic. In this paper, we conduct a systematic study to compare different deep learning models, including CNN, GRUs, and LSTM networks, as well as their hybrid architectures such as CNN- GRU and CNN-LSTM on real-world NB-IoT dataset with benign traffic traces mixed up against targeted attacks from Mirai/Gafgyt botnets. The CNN-LSTM hybrid model demonstrated significant performance in IoT intrusion detection, achieving accuracy rates of 94.7%, precision of 94.6%, recall of 94.7%, and F1-score of 94.6%.