Computational model of multi-source data fusion for submarine cable threat detection

Abstract

Submarine cable monitoring relies primarily on the automatic identification system, which is vulnerable to spoofing attacks that can cause early-warning failure. This paper proposes Adaptive Spatio-temporal Graph Attention Network with Evidence Fusion, a multi-source data fusion model for submarine cable threat detection that couples spatio-temporal graph representation learning with adaptive Dempster–Shafer evidence discounting in one end-to-end framework. Cable topology, automatic identification system trajectories, and underwater acoustic data are aligned onto a unified spatio-temporal graph, from which cross-domain features are extracted through graph convolution and attention. At the decision level, an adaptive conflict-discounting mechanism prioritizes comparatively tamper-resistant acoustic evidence when digital and physical evidence conflict, thereby resisting spoofing. On a real ocean dataset, the model attains a balanced accuracy of 0.975 and a recall of 0.951, and maintains a harmonic mean of precision and recall of 0.910 under 30% spoofing attack. The framework is scalable for securing marine critical infrastructure.