End-to-end deep learning for smart maritime threat detection: an AE-CNN-LSTM-based approach.

Smart maritime operations face growing cyber risks due to the proliferation of IoT-enabled sensors, navigation units, and communication links. To improve detection fidelity under these conditions, we present a hybrid Autoencoder-Convolutional Neural Network-Long Short-Term Memory (AE-CNN-LSTM) based framework that unifies unsupervised reconstruction signals with spatio-temporal feature learning for intrusion detection in marine cyber-physical networks. The model is trained and evaluated on a KDDCup99-based benchmark adapted to simulated maritime scenarios and supports both binary and multiclass classification. In the binary setting, the system attains 99.8% accuracy; in the multiclass setting it demonstrates consistently strong performance across precision, recall, F1-score, and AUC, with minority-class behavior analyzed via confusion matrices and threshold sensitivity. Reconstruction errors (MAE/MSE) provide an auxiliary anomaly cue that aids triage. In this study the results are compared with representative deep-learning and transformer baselines, the proposed model yields competitive to superior results while remaining suitable for real-time deployment in smart ports, autonomous vessels, and underwater sensor networks. We also discuss practical constraints-such as dataset realism and class imbalance-to contextualize applicability in operational environments.