Edge-AI Enabled Intrusion Detection Framework for Ultra-LowLatency IoT Networks in Smart Cities

Abstract

Internet of Things (IoT) infrastructures are highly dependent on smart cities to be able to handle basic amenities like intelligent transportation and smart utilities, community security, and environmental management. Nevertheless, the spread of IoT devices with limited resources opens networks to cyberattacks that may interfere with important services. Conventional cloud-based intrusion detection systems (IDS) add spoilage and bandwidth and limited scalability. Recent developments in edge computing and minimalistic artificial intelligence models promise the solutions to real-time threat mitigation. This paper will propose a prototype of an Edge-AI based intrusion detection system, which can be used in ultra-low-latency IoT networks used in smart city architectures. A simulated and real-world dataset of IoT traffic was trained on lightweight machinelearning models, such as MobileNet-V3 embedded classifiers, optimized Random Forest, and quantized neural networks to be deployed on the edge. The paper tested the performance of the model-based edge devices, including the NVIDIA Jetson nano, Google Coral TPU, and ARM-based microcontrollers. The proposed structure proved to have an average detection rate of 94.2 with latency of less than 12ms and a 38 percent energy saving over cloud based IDS. The experimental findings forwarded that the implementation of AI on the network edge can significantly accelerate the time of threat response, reduce the loss of packets, and increase the general system resilience. The results reveal that Edge-AI models can be successfully applied to achieve ultra-lowlatency IoT networks, which can be scaled and autonomously used in the cybersecurity of smart cities. It is suggested that the large-scale deployments and federated learning should be integrated in the future to increase robustness or privacy.