Hyperparameter Tuning of Convolution Neural Network for Intrusion Detection Systems using Particle Swarm Optimization

A research challenge is to investigate multi-class classification methods that can successfully identify and categorize intrusions in network traffic problems. Convolutional Neural Networks (CNN) have been shown to successfully detect and classify multiple types of network anomalies. However, to automatically leverage its performance, the hyperparameters of the CNN usually need to be tuned. We study the use of three existing Particle Swarm Optimization (PSO) variants: hybrid cooperative PSO (CPSO-HK), quantum PSO (QPSO), adaptive PSO (APSO), and introduce a new variant QAPSO, for the purpose of tuning hyperparameters of a 1D-CNN model. This work examines the effects this tuning has on improving the model’s ability to classify multiple Botnet intrusion attacks. The QAPSO variant is shown to converge significantly faster than all other variants when tuning. The results indicate that the structure size of the model is influenced by the PSO variant tuning. It is also found that during training and testing, models tuned with QAPSO and QPSO produce more consistent results and are able to outperform models tuned with CPSO-HK and APSO.