Lightweight Intrusion Detection Based on Hybrid Feature Selection Machine Learning

Industrial control security is the lifeblood of a country and the core component of production and life. It targets the issues of diverse attack methods and unbalanced training sample data in traditional industrial control systems. This paper proposes a novel hybrid machine-learning model that integrates PSO-CNN-Attention-BiLSTM. SMOTE algorithm is used to solve the problem of small samples and unbalanced attacks in training and generate missing sample data. The backbone network employs a CNN-Attention-BiLSTM hybrid model for training, extracting feature characteristics from sample data. The backbone network considers the time series relationship in the input sample and the degree of spatial feature relationship inside the sample. It uses an attention mechanism to extract relationship features within samples. Finally, it uses the classification module to classify the prediction results. Through experimental tests, the model has a classification accuracy of 98.18% and an F1 score of 0.9809 on the intrusion detection data set of an industrial production PLC device in Hebei Province. Compared with the conventional intrusion detection model, this method has a higher classification accuracy and better performance than the traditional method.