Intrusion Detection With Machine and Deep Learning: A Comparative Study on CICIDS‐2017 and NSL‐KDD

With the rapid advancement of technology, the complexity and frequency of cyberattacks are increasing, necessitating advanced defensive measures. Intrusion detection systems (IDS) play a critical role in mitigating these threats. This study investigates the effectiveness of diverse algorithmic architectures—including Machine Learning (ML) (Random Forest, Decision Tree [DT], KNN), Deep Learning (DL) (CNN, DNN, LSTM, BiLSTM, GRU, BiGRU), and Hybrid models (CNN+BiLSTM, GRU+Random Forest)—using the NSL‐KDD and CICIDS‐2017 datasets as established benchmark datasets. To optimize performance, Principal Component Analysis (PCA) and SelectKBest + Mutual Information (SKB+MI) were applied for feature selection. Experimental results show that while the DT algorithm achieved a peak accuracy of 99% on the CICIDS‐2017 dataset, it demonstrated sensitivity to dataset‐specific artifacts. In contrast, the GRU+Random Forest hybrid model demonstrated favorable performance and operational efficiency, achieving a testing latency of 0.03 s. This research provides a comparative framework for examining the trade‐off between detection precision and computational overhead, offering insights into cross‐dataset generalizability and potential deployment considerations for high‐throughput network environments.