Detection of Ongoing GPS Spoofing Using A Convolutional Neural Network

Abstract

Deliberate GPS interference by broadcasting fake signals, so-called spoofing, can compromise critical infrastructure that relies on location and timing services. This work develops a dual-branch convolutional network (CNN) that automatically detects spoofing attacks during the acquisition stage of a GPS receiver. The receiver’s 2-D acquisition map (a single-channel image) is processed both locally (region-ofinterest, ROI) and globally to capture spoofing cues. The CNN was trained and evaluated on MATLAB-simulated recordings as well as on two publicly available datasets, the Oak Ridge Spoofing and Interference Test Battery (OAKBAT) and the Finnish Geospatial Research Institute (FGI) repository, which together comprise 11 GPS recordings containing spoofing scenarios. Results show reliable detection when training and testing on data from the same domain, achieving balanced-accuracy scores of 99 % on targeted scenarios. Cross-domain generalization to datasets with different parameters decreases noticeably, with performance dropping to nearrandom- guessing levels. The study also provides visual illustrations of authentic and spoofed scenarios. The total average inference time of the solution is 22-23 ms per tracked satellite, indicating practical applicability. These findings suggest that a dual-branch CNN operating in the acquisition stage is a viable method for spoofing detection.