Deep learning has revolutionised pattern recognition by integrating feature extraction and classification into end-to-end models, eliminating the need for manual feature engineering. This survey paper reviews recent advancements in deep learning–based anomaly detection for network intrusion detection systems (NIDS). The review focuses on three major classes of models: convolutional neural networks (CNNs), recurrent architectures (particularly LSTMs), and autoencoders. Each class demonstrates distinct strengths in capturing spatial, temporal, and latent representations of network traffic. The paper synthesises key studies applying these models to anomaly-based intrusion detection, compares reported performance across commonly used benchmark datasets, and discusses their effectiveness in detecting previously unseen attacks. Finally, it highlights ongoing challenges such as class imbalance and concept drift, and outlines future research directions, including adversarial training and online adaptation, to enhance deep-learning-driven NIDS.
Johnson et al. (Wed,) studied this question.