A shipborne GNSS spoofing detection algorithm based on temporally constrained prior-association

Abstract Ship navigation relies on the global navigation satellite system (GNSS), but civil GNSS signals are vulnerable to spoofing attacks that can induce erroneous positioning results. Existing inertial- and sensor-based methods usually describe the consistency between GNSS information and measurements from non-GNSS sensors through direct motion comparison, filter innovations, or test statistics derived from acceleration constraints, but they do not fully characterize the temporal association changes in the motion inconsistency caused by spoofing attacks in maritime scenarios. We propose a shipborne GNSS spoofing detection algorithm based on temporally constrained prior-association. Orientation is estimated by fusing an inertial measurement unit (IMU) and a gyrocompass, with the observation noise covariance of the normalized gravity vector adaptively adjusted according to the gravity magnitude deviation. The IMU acceleration is transformed into the navigation frame, and the residual between the second-order-differenced GNSS acceleration and the IMU acceleration is used as the input sequence. A transformer detection framework is then constructed by introducing a prior-association weight matrix jointly constrained by temporal distance and temporal position into multi-head attention. The symmetrized Kullback-Leibler (KL) divergence between the series-association and prior-association matrices is combined with reconstruction error to form the anomaly score. Ablation experiments on real ship navigation data with injected spoofing attacks show that temporally constrained prior-association provides the main improvement in detection performance, with an additional gain from adaptive observation noise covariance. On test subset 2, comparison results pooled across all attack types and their configured intensities yield an aggregate missed detection probability of 4.92% at an aggregate false alarm probability of 5%. The aggregate detection rate within 30 s reaches 91.62% at an aggregate false alarm probability of 1%. The computation-time experiment indicates a higher computational cost.