Accelerometer Aided GNSS Spoofing Detection Using Wavelet Based Time-Frequency Analysis

Abstract

Global Navigation Satellite System (GNSS) receivers are widely used in safety-critical applications such as autonomous driving and aviation, where accuracy and integrity are essential. However, civilian GNSS signals are weak and unauthenticated, which makes them vulnerable to spoofing: the broadcast of falsified GNSS signals that causes a receiver to report an incorrect position, velocity, or time. Previous approaches compare GNSS-derived and inertial measurement unit (IMU) accelerations using statistical hypothesis tests, but these tests have limited sensitivity to attacks that preserve the marginal signal statistics. In this thesis, a spoofing detection framework based on the wavelet cross-coherence between the IMU- and GNSS-derived acceleration signals is described. The continuous wavelet transform is computed independently on each signal, and the magnitude-squared coherence and cross-phase are evaluated in the time-frequency plane. The key insight is that even when spoofing leaves the marginal residual statistics almost unchanged, it breaks the time-frequency agreement between the two sensors. A per-frequency window-mean (PFWM) detector is developed: it averages the coherence and phase within each frequency bin in the [0.1-0.5] Hz band over sliding windows and fits a Beta distribution to clean modelling data to set the per frequency bin thresholds. The detector is evaluated on a 1381 s real-vehicle dataset collected across three independent drives in Blacksburg, Virginia (540 s of modelling data and an 841 s test drive), under a time-delay replay attack in which the adversary captures the live GNSS signal and rebroadcasts it with a controlled time delay at higher power, causing the victim receiver to lock onto a delayed but internally consistent navigation solution. The proposed wavelet detector achieves 100% detection and 0% clean false-alarm rate for replay delays up to 8 s, whereas the classical z and chi^2 tests remain below 70% under the same conditions. However, it is also shown that wavelet coherence is structurally blind to constant-bias spoofing due to the zero-mean (admissibility) property of the wavelet, and effectively blind to ramp spoofing because the energy of a slowly growing ramp is concentrated at very low frequencies, well below the detector's analysis band. In both cases, classical statistical tests are able to detect the attack.