OSPXAI: An Explainable and Quantifiable Deep Learning Method for the Detection and Identification of Serial Radiation Leakage Signals

Electromagnetic radiation leakage during serial communication poses critical security vulnerabilities in data transmission. This paper presents OSPXAI, a deep learning methodology that integrates feature extraction and model interpretability for serial radiation leakage signal detection. The architecture incorporates three primary contributions: 1) an Offset Scaling Pooling Layer (OSPLayer) that implements adaptive feature extraction through offsetting, scaling, and rotational positional coding mechanisms; 2) an enhanced Smooth Gradient-weighted Class Activation Mapping (Smooth Grad-CAM++) that generates interpretable feature visualization maps; and 3) a novel Signal Overlap Index (SOI) metric that quantifies the intersection-over-union between predicted and actual signal regions. Experimental validation on RS-232 serial signals demonstrates that OSPXAI achieves 99.41% classification accuracy in binary detection tasks. The model exhibits robust generalization across multi-class recognition tasks with varying baud rates (9600-115200), maintaining a more stable performance under complex noise conditions. Quantitative evaluation through SOI validates the model’s feature localization capabilities, while Smooth Grad-CAM++ visualization provides interpretable insights into the decision-making process. The proposed methodology advances both the accuracy and explainability of electromagnetic leakage detection, offering practical value for communication security applications.