Integrating prompt-based zero-shot learning into Bayesian network for maritime cybersecurity risk analysis

The increasing digitalisation of maritime ecosystems, including vessels, ports, offshore facilities, and shore-based organisations, has expanded the attack surface of maritime operations and increased exposure to cyber threats. Cyber incidents affecting these interconnected systems may propagate across global logistics networks, highlighting the need for systematic cybersecurity risk analysis in maritime transport. However, existing studies often rely on expert judgement or limited datasets, which restrict scalability and hinder comprehensive identification of Risk Influential Factors (RIFs). This study develops a data-driven framework that integrates Large Language Model (LLM)-assisted text classification with Bayesian Network (BN) modelling to analyse maritime cybersecurity risks. A structured taxonomy of RIFs is defined to capture temporal, regional, operational, and technical attributes of cyber incidents. LLM-based prompt engineering is then employed to automatically extract these factors from unstructured incident reports. The reliability of the automated classification process is evaluated through manual annotation on a sampled subset of incident reports and comparative experiments across multiple LLMs. The BN model is further assessed using repeated cross-validation to evaluate predictive performance and robustness. The results reveal key dependencies among cyber incident characteristics and consequence types, providing probabilistic insights into maritime cyber risk patterns. The proposed framework offers a scalable approach for analysing cyber risks in maritime transport systems and supports risk-informed cybersecurity management.