A New Method to Improve the CNN Configuration for IoT Attack Detection Problem Based on the Genetic Algorithm and Multi-Objective Approach

This paper proposes a new method to enhance the configuration of convolutional neural network for IoT attack detection based on genetic algorithm and multi-objective approaches. According to the multi-objective approach, the paper builds component objective functions based on accuracy, precision and recall measures; the global objective function is built on three component objective functions with corresponding weights; which can be adjusted to suit different problem domains. The global objective function serves as the fitness function in the genetic algorithm; it is used to select promising chromosomes after each generation. Each configuration of the CNN will be represented as a chromosome. The search space is all possible configurations of the CNN. In each generation, two chromosomes will perform crossover to create two offspring chromosomes - which are two new configurations; one chromosome will perform mutation to create a new chromosome - a new network configuration. The crossover and mutation rates are optional in the problem. Our proposed method is tested on the Edge-IIoTset dataset, achieving the most balanced network configuration between false positive, false negative and accuracy measures; in which the achieved global objective function value is 95.11%, the achieved improvement rate is 0.07%; the achieved accuracy is 94.81%, the improvement rate of the accuracy is 0.04%.