A newly developed deep learning framework combining convolutional networks, bidirectional temporal modeling, and global attention mechanisms can detect stealthy cyberattacks against maritime power systems with 97.71 percent accuracy while reconstructing compromised sensor telemetry with a determination coefficient of 0.9686. Published in Intelligent Marine Technology and Systems, the study introduces an integrated architecture called CNN-BiLSTM-GAM to defend shipboard cyber-physical systems against false data injection attacks. Modern commercial vessels increasingly rely on interconnected digital networks to regulate electrical distribution, yet malicious actors can introduce falsified sensor readings that evade standard defenses. In experimental evaluations, the framework achieved an F1-score of 0.9548 in attack detection, significantly outperforming conventional baseline models. By simultaneously identifying subtle disruptions and restoring corrupted measurements, the model establishes a data-driven defense mechanism designed to maintain operational resilience across automated maritime platforms.
This development comes amid accelerating digital transformation across the maritime shipping industry, as vessels transition from isolated mechanical units into interconnected cyber-physical networks. Critical operational technology systems governing propulsion, navigation, and power control are now linked in real time to shore control centers through ship-to-shore communications links. Although connectivity improves operational efficiency, it drastically expands the attack surface exposed to remote intrusion. Maritime cyber incidents rose by 300 percent by 2020, prompting the International Maritime Organization to mandate cyber risk management in ship safety frameworks. Real-world incidents demonstrate the severity of these vulnerabilities; in 2017, a container vessel experienced a remote navigation compromise lasting ten hours during an attempted hijacking, while the NotPetya attack disrupted global operations for A.P. Moller-Maersk, inflicting financial losses approaching 300 million dollars.
At the center of maritime operational integrity is the shipboard power system, which sustains critical navigational, propulsion, and auxiliary machinery. Modern vessels increasingly structure electrical systems as complex shipboard microgrids integrating generation units, energy storage devices, and heavy propulsion loads. To ensure electrical stability, the vessel energy management system relies on state estimation as its primary cognitive mechanism. State estimation processes noisy sensor data to calculate the most credible real-time operational state of the electrical grid. This calculated state serves as the definitive source of truth for downstream automated decisions, including economic power dispatch, protective relay settings, equipment fault prediction, and emergency load shedding routines. If sensor measurements are compromised, state estimation generates distorted calculations, potentially causing improper control responses that risk widespread electrical failures, severe machinery damage, or complete vessel blackout.
Among maritime cybersecurity threats, false data injection attacks present an exceptionally grave danger because of their ability to bypass traditional surveillance systems. Classical supervisory control systems rely on bad data detection mechanisms embedded within weighted least squares state estimation, checking measurement residuals representing differences between observed sensor readings and model-derived estimates. However, adversaries possessing knowledge of the system Jacobian matrix can construct coordinated attack vectors that align with the physical network model. By injecting structured bias across multiple correlated sensor channels during transmission, attackers manipulate the calculated state while keeping post-attack measurement residuals identical to uncorrupted conditions. Because the bad data detector observes residuals that remain below established alarm thresholds, the attack remains completely hidden. Consequently, adversaries can subtly alter vessel physical behavior without triggering conventional residual-based security alarms.
To overcome blind spots in conventional residual testing, the research team designed an integrated deep learning framework uniting attack detection and data reconstruction into a continuous pipeline. Rather than separating threat identification from mitigation, the CNN-BiLSTM-GAM architecture establishes a closed-loop system built upon a shared feature extraction backbone. In this dual-model strategy, multivariate supervisory control and data acquisition telemetry is continuously screened for malicious anomalies. When the classification module identifies compromised sensor channels, the system automatically triggers an integrated recovery module to reconstruct authentic values using learned spatiotemporal contextual dependencies. These reconstructed measurements immediately replace compromised readings within the telemetry stream, providing a purified data feed to the state estimator. This autonomous closed-loop mechanism preserves system observability, prevents biased state estimates, and ensures operational continuity during cyber disruptions without manual operator intervention.
The technical structure of the framework utilizes a multi-stage deep learning pipeline designed to untangle complex maritime sensor dynamics. The model first deploys a convolutional neural network featuring a kernel size of three and thirty-two output channels to capture local cross-sensor spatial correlations across incoming telemetry streams. Sliding convolutional kernels along the temporal axis extracts spatial dependencies among distributed electrical sensors, while pooling layers compress dimensionality and preserve dominant spatial patterns. The resulting spatial feature sequence is passed to a bidirectional long short-term memory network containing thirty-two hidden units per direction. While standard recurrent models only track historical states, the bidirectional architecture processes forward and backward temporal sequences simultaneously. This bidirectional tracking allows the network to capture complex temporal dynamics, including dynamic load fluctuations and predictable power declines observed during evening operating cycles.
To enhance operational stability in noisy maritime environments where environmental disturbances can mask stealthy attacks, the researchers integrated a global attention mechanism following the recurrent modeling layer. Because spatial and temporal features contribute unequally to attack identification, the attention mechanism dynamically recalibrates feature representations across channel and temporal dimensions. The channel attention module aggregates spatial information across channels using global pooling, subsequently employing a compact multi-layer perceptron to compute channel-specific weights that emphasize informative feature pathways. Following channel refinement, a spatial attention module examines the temporal dimension by pooling along channel axes and applying convolution to construct a temporal attention map. This dual-stage recalibration highlights subtle anomaly signatures induced by attacks while suppressing ambient noise and routine fluctuations, generating an enriched sixty-four-dimensional feature representation for downstream decision-making.
The architecture handles threat identification and data reconstruction by mapping attention-refined feature representations into distinct operational tasks. Attack detection is formulated as a multi-label binary classification task, where a dense output layer transforms enhanced features into probability values aligned with individual sensor channels. Each probability reflects the likelihood that a specific measurement was compromised, evaluated against a fixed threshold of 0.5 to balance missed detections against false alarms, and trained using binary cross-entropy loss. Simultaneously, data recovery is formulated as a time-series regression task that reconstructs compromised sensor values from uncorrupted contextual relationships. The recovery network is trained on normal operational data using mean squared error loss. When an attack is flagged, the model reconstructs authentic measurement points, applies inverse normalization to restore original physical units, and outputs a purified data stream directly to the vessel state estimator.
The research team evaluated the framework using empirical data collected from an offshore photovoltaic plant in Europe, combined with simulated false data injection scenarios created in MATLAB R2024a. The complete experimental dataset contained 8,737 multivariate time-series records, with anomalous samples comprising 25 percent of the total volume. Computational simulations were conducted using Python 3.12 in PyCharm Community Edition 2024.3.1 on a workstation equipped with an AMD R9 7945HX central processing unit and an NVIDIA RTX 4060 graphics processing unit. The data were partitioned into training, validation, and testing subsets according to an 8:1:1 ratio, with min-max normalization applied to harmonize distinct sensor units.
Systematic ablation experiments confirmed that each component within the hybrid architecture contributed significantly to overall detection capability. A standalone LSTM baseline achieved an accuracy of 0.9577 plus or minus 0.0032. Incorporating convolutional layers in the CNN-LSTM model increased accuracy to 0.9680 plus or minus 0.0024, representing a 1.08 percent gain. Upgrading to bidirectional temporal modeling in the CNN-BiLSTM configuration further improved accuracy to 0.9714 plus or minus 0.0019, an increase of 1.43 percent over the standalone LSTM. The complete CNN-BiLSTM-GAM framework achieved superior results, reaching an accuracy of 0.9771 plus or minus 0.0014, an F1-score of 0.9548 plus or minus 0.0015, precision of 0.9856 plus or minus 0.0019, and recall of 0.9635 plus or minus 0.0017. Robustness evaluations across disturbance levels demonstrated detection accuracies of 0.9325 under low-intensity stealth attacks, 0.9771 under medium intensity, and 0.9943 under high intensity, requiring 17.348 seconds of training and 0.155 seconds for test inference.
In localization benchmarks across eight variables—device temperature, ambient temperature, atmospheric pressure, relative humidity, global horizontal irradiance, direct normal irradiance, diffuse horizontal irradiance, and electrical power—the framework attained the highest accuracy for every metric, outperforming both transformer and graph neural network baselines. In data recovery evaluations, the CNN-BiLSTM-GAM module significantly exceeded baseline performance, achieving a determination coefficient of 0.9686 plus or minus 0.0014 compared to 0.9219 for a multi-layer perceptron and 0.9503 for a transformer model. Error analysis across 100 consecutive sampling points indicated that reconstructed power curves closely tracked ground-truth profiles, accurately capturing steep transitions and peak values. Statistical analysis of prediction errors confirmed that the framework produced residual distributions tightly centered near zero without systematic bias across varying load ranges, requiring 65.855 seconds of training and 0.187 seconds for inference execution.
Computational profiling indicated that the architecture maintains a low memory footprint and an inference latency of approximately 40 milliseconds per sample, confirming its feasibility for deployment on marine computing platforms. Nevertheless, the study highlighted clear experimental limitations. The authors acknowledged that an offshore photovoltaic dataset does not reflect the full operational heterogeneity of a true shipboard microgrid, which incorporates diesel generators, gas turbines, and dynamic propulsion loads. Additionally, simulated attack vectors followed standard stealth assumptions rather than specialized maritime intrusion patterns. To transition the framework toward active service, future research must validate the system on hardware-in-the-loop testbeds connecting real-time simulators with physical protection relays. Subsequent efforts will also explore federated learning to facilitate multi-vessel collaborative model training without compromising proprietary data, alongside network pruning, model quantization, and domain-adaptive transfer learning to streamline deployment across diverse commercial fleets.
Subject of Research: Deep learning defense against false data injection attacks in maritime cyber-physical power systems
Article Title: Intelligent spatiotemporal-fused deep learning resilient framework for maritime cyber-physical systems
Article References: Wang, X., Han, W., Luo, X., & Guan, X. (2026). Intelligent spatiotemporal-fused deep learning resilient framework for maritime cyber-physical systems. Intelligent Marine Technology and Systems, 4(1), Article 20. https://doi.org/10.1007/s44295-026-00109-3
Image Credits: AI Generated
DOI: 10.1007/s44295-026-00109-3
Keywords: maritime cyber-physical systems, false data injection attacks, shipboard power systems, state estimation, CNN-BiLSTM-GAM, bad data detection, shipboard microgrids, spatiotemporal modeling, global attention mechanism, data recovery, sensor data reconstruction, bidirectional LSTM, convolutional neural networks, maritime cybersecurity, anomaly detection, SCADA systems, offshore photovoltaic power
Cite Scienmag News
Blake Davidson. (November 18, 2026). Deep Learning Framework Detects Maritime Power Cyberattacks With 97.7 Percent Accuracy. Scienmag. https://scienmag.com/deep-learning-framework-detects-maritime-power-cyberattacks-with-97-7-percent-accuracy/
Blake Davidson. "Deep Learning Framework Detects Maritime Power Cyberattacks With 97.7 Percent Accuracy." Scienmag, 18 November 2026, https://scienmag.com/deep-learning-framework-detects-maritime-power-cyberattacks-with-97-7-percent-accuracy/. Accessed 22 November 2026.
Blake Davidson. "Deep Learning Framework Detects Maritime Power Cyberattacks With 97.7 Percent Accuracy." Scienmag. November 18, 2026. https://scienmag.com/deep-learning-framework-detects-maritime-power-cyberattacks-with-97-7-percent-accuracy/








