Abstract: The emerging Internet of Things (IoT) 5G smart cities are experiencing radical changes in transportation, utilities and energy by real-time interconnection, automation, and data sharing. Gone are the days when we could just protect these vulnerable assets on the physical network layer. As things are increasingly open for attack, we can no longer afford being under protected. To mitigate the above problems, this paper puts forward a lightweight scalable Hybrid Deep Learning-Based Security Architecture designed to secure IoT systems in 5G Smart City environments. The proposed model leverages DNN for the spatial feature extraction and LSTM networks to capture the temporal relationships in network traffic. Transfer learning is leveraged in the architecture to improve flexibility and detection accuracy by reutilizing prior learned knowledge and only retraining to account for new threat vectors. The model is trained and tested using benchmark datasets that cover diverse real-world attack landscapes like CICIDS-2018 and UNSW-NB15. According to the results of the experiments, the filtered model yields a low false positive rate of 1.08%, the AUC-ROC is 0.987 and exhibits good performance measures, such as an accuracy of 99.74%, precision of 98.21%, recall of 99.89%, and F1-score of 98.05%. These results are evidence of the model's real-time intrusion detection capability, making the framework a feasible and effective technique for securing the 5G-enabled IoT infrastructures of smart cities.

Key words: 5G smart cities, IoT security, hybrid deep learning, intrusion detection system, transfer learning