Identificación de ataques de denegación de servicio distribuido (DDoS) mediante la integración de algoritmos de aprendizaje automático y arquitecturas de redes neuronales artificiales.
Víctor Alfonso Guzmán-Brand, Laura Esperanza Gélvez-García · Revista de Ingeniería Matemáticas y Ciencias de la Información · 2025
Objective: To identify distributed denial of service (DDoS) attacks by integrating machine learning algorithms and artificial neural network architectures. Methodology: To structure the data analysis, the Knowledge Discovery Data (KDD) technique is used. This approach allows examining large volumes of information of various types, with the objective of identifying patterns, correlations and producing valuable information. As for the data set, the CIC-DDoS2019 dataset developed by the Canadian Cybersecurity Institute is used. Results: When training and evaluating the different algorithms, it was observed that the models based on decision trees, such as Random Forest and XGBoost, stood out for achieving the best results in terms of accuracy and efficiency. On the other hand, in the analysis of the performance of the neural networks, the Closed Stream Units (GRU) stood out by obtaining the best results in accuracy and precision. This performance suggests that GRUs achieve an optimal balance between predictive ability and minimization of false positives and negatives. Discussion: In the comparison between traditional machine learning models and neural networks for DDoS attack detection, it is observed that algorithms such as XGBoost and Random Forest offer similar or superior performance in terms of accuracy and also exhibit significantly shorter execution times. On the other hand, neural networks such as GRU and RNN achieve high accuracy, but with a high computational cost. Conclusions: XGBoost, demonstrated an optimal balance between accuracy (F1-score: 0.9992) and speed (11.47s), positioning itself as the most viable alternative for real-time implementations. In the field of neural networks, Gated Stream Units (GCU) obtained the best performance (accuracy: 0.9992; F1-score: 0.9992), given the ability to process temporal dependencies and reduce false positives.