Redes Neurais Artificiais combinadas a algoritmos genéticos para detecção de intrusões em rede de computadores

Abstract

This work presents a methodology for intrusion detection in computer networks based on the integration of Genetic Algorithms (GAs) and Convolutional Neural Networks (CNNs). The proposal simultaneously contemplates the automatic selection of attributes that best characterize network traffic and the adjustment of CNN hyperparameters, with the aim of maximizing performance metrics and minimizing inference latency. Using the CIC-IDS2018 dataset, the experiments showed that the attribute selection GA was able to identify a subset of 20 variables that preserves the model's discriminative capacity, and that the hyperparameter adjustment GA, in turn, produced a CNN configuration with slightly superior performance, achieving an accuracy of 99.78%, an F1-score of 99.66%, and a recall of 99.82%, with an average latency of only 0.0529 ms per sample. Qualitative analysis demonstrated that the selected attributes are directly relevant in characterizing traffic patterns, while the optimized hyperparameters resulted in better separability between benign and malicious flows. Confusion matrices indicated a low incidence of false positives and false negatives, reinforcing the robustness of the approach. The results obtained surpassed those of related studies in the literature, validating the effectiveness of the proposal.