Network Intrusion Detection Systems (NIDS) are a critical line of defence against the growing volume and sophistication of cyberattacks targeting modern networked infrastructure. While deep learning has substantially advanced anomaly and misuse detection, a systematic comparison of representative architectures under a unified evaluation protocol across heterogeneous traffic domains remains limited. This paper presents a comparative study of five deep learning architectures—Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), Gated Recurrent Units (GRU), Autoencoders (AE), and Transformer-based classifiers—evaluated on three widely used benchmark datasets: NSL-KDD, CICIDS2017, and UNSW-NB15. Experiments are conducted under a common preprocessing, cross-validation, and metric-reporting protocol to ensure fair comparison. Results show that Transformer-based models achieve the highest detection accuracy and AUC-ROC overall, recurrent architectures (LSTM/GRU) offer the strongest sequence-modelling advantage on flow-based traffic, and Autoencoders provide the most computationally efficient route to unsupervised anomaly detection at a moderate accuracy cost. The findings offer practical guidance for selecting deep learning architectures in intrusion detection deployments under differing accuracy, latency, and resource constraints.

Keywords — Intrusion Detection, Deep Learning, Network Security, Convolutional Neural Network, LSTM, GRU, Autoencoder, Transformer, Anomaly Detection, CICIDS2017