A Novel Approach for 1D-CNN Hyperparameter Optimization in IoT Attack Detection using PSO

This study proposes a hyperparameter optimization method for one-dimensional convolutional neural network using the Particle Swarm Optimization algorithm based on a Pareto multi-objective approach to improve the performance of IoT attack detection systems. Specifically, this study enhances the PSO algorithm by introducing an automatic termination criterion for optimization loops and proposes an early stopping mechanism, along with the optimization of the early stopping patience during the 1D-CNN model training process, thereby reducing computational costs and aligning with the resource-constrained hardware conditions of IoT. Additionally, a multi-objective optimization function is developed to balance detection performance and resource efficiency by combining validation accuracy with the 1D-CNN's execution time. The proposed method is evaluated on the Edge-IIoTset dataset. Experimental results demonstrate that the optimized model reduces execution time by 48-63% compared to the baseline model while maintaining high accuracy (over 94%). This research not only provides a practical solution for IoT security but also pioneers a novel approach to integrating evolutionary algorithms into adaptive deep learning systems and introduces a flexible method for hardware-constrained devices.