DAS-Based Perimeter Intrusion Detection Using 2D-CNN with SK Attention Mechanism

Distributed fiber optic sensing technology has been extensively applied in the field of perimeter security. The distributed acoustic sensing (DAS) system driven by deep learning architecture has significantly the efficiency and accuracy of intrusion detection. However, achieving high-precision recognition of events with subtle feature differences under complex environmental conditions remains a significant challenge. To address this issue, this study proposes a novel model architecture that combines a conventional two-dimensional convolutional neural network (2D-CNN) with a selective kernel (SK) attention module. By dynamically and adaptively adjusting feature weights, the model enhances the recognition accuracy of interference events in complex environments. Experimental results demonstrate that the 2D-CNN-SKAttention model effectively identifies different types of interference events, achieving an accuracy of 99.3%, which represents a 3.8% improvement over the standard 2D-CNN model. Furthermore, rigorous analysis of classification performance across varying signal-to-noise ratio (SNR) conditions reveals sustained accuracy exceeding 98% even at an SNR of -2 dB. This result underscores the robustness and reliability of the model in high-noise scenarios, confirming its practical applicability for perimeter security systems operating under adverse environmental conditions.