CNN-PSO-KELM: A Deep Learning Intrusion Detection Model for Imbalanced IoT Data

Both the real network environment of IoT devices and the most of publicly available datasets suffer from the problem of sample imbalance. Traditional intrusion detection models fail to effectively identify minority samples. Therefore, this paper proposes a CNN-PSO-KELM model aimed at enhancing the model’s detection capabilities in imbalanced data, particularly, for minority samples. This model employs an improved dual-layer convolutional neural network (CNN) for feature extraction, followed by label classification using kernel extreme learning machine (KELM). It achieves the combination of the generalization ability of CNN and the high learning ability of KELM. To address the performance degradation caused by the random initialization of KELM parameters, particle swarm optimization (PSO) is introduced to assist in parameter acquisition for KELM. The effectiveness of this improvement has been verified on three IoT datasets: NSL-KDD, CIC-IDS2017, and Bot-IoT. Compared to traditional deep learning algorithms, CNN-PSO-KELM significantly improves the detection accuracy for imbalanced data.