A Novel MDI-Pruning for Low-Latency Deep Learning-Based IoT IDS on Raspberry Pi

Several studies have shown that deep learning techniques, such as deep neural networks (DNNs) using fully connected layers, can detect malicious traffic with high performance. However, there are challenges in deploying deep learning on Internet of Things (IoT) devices, such as the Raspberry Pi, namely its limited computing power and energy constraints. The purpose of this study is to observe the effect of applying pruning to deep learning in an intrusion detection system (IDS) for the Internet of Things (IoT) using Raspberry Pi. We use the RT-IoT 2022 dataset from Kaggle. In pre-processing, we conduct several methods: cyber-attack clustering, mean decrease in impurity (MDI) feature selection, label encoding, and standardization. The next step is to design and optimize deep learning performance for IDS. Then, we design and optimize pruning for deep learning, considering both model size and accuracy. In the last step, we deploy both models on a Raspberry Pi and compare their latency. The test results show that the original DNN method is 30.3 kB. Other methods are smaller than the DNN method. However, the smallest is the DNN+MDI+pruning method, which is 7.8 kB. With this value, the CR of DNN+MDI+pruning is 3.9×. Thus, the DNN+MDI+pruning method becomes the method with the highest CR and the lowest accuracy loss compared to other methods. DNN+MDI+pruning is also a model with a significantly lower average latency than the other three models, at 0.70 seconds.