A Lightweight CoDeV–HyPerion‐T Deep Learning Framework for Threat Detection in Underwater Wireless Sensor Networks (UWSNs)

ABSTRACT

The use of underwater wireless sensor networks (UWSNs) has become a major security concern due to their increased use in environmental monitoring, military surveillance, and offshore applications. The networks are highly susceptible to various cyber threats due to their dynamic topology, limited energy resources, and unreliable communication networks. The existing intrusion detection techniques tend to exhibit low detection rates, high false alarms, and high complexity, making them inapplicable to real‐time and resource‐constrained UWSN systems. Moreover, many conventional and deep learning‐based models cannot effectively model feature interactions and emerging attack patterns in heterogeneous IoTs and underwater environments. To address these weaknesses, this paper presents a new hybrid intrusion detection model, which integrates CoDeV based feature extraction algorithm and the HyPerion‐T model. To capture both the spatial and temporal characteristics of network traffic and leverage a transformer‐based structure to enhance intrusion detection by exploiting adaptive attention, the article presented here was based on sophisticated feature description and dependency modeling. These combinations will ensure greater accuracy, strength, and efficiency in the calculations. The results of the proposed methodology are tested on the benchmark datasets, i.e., WSN‐DS, IoT‐23, and N‐BaIoT. Among all models, CoDeV+HyPerion‐T has reached the highest average accuracy (98.96%) with an accuracy of more than 98% and recall of more than 98% and also achieved the lowest average inference time (9.5 ms), demonstrating great real‐time performance. The inference energy is also measured to be 3.42 mJ for per inference, achieving up to 50.1% energy reduction compared with basis approach models.