RDPA: Dual-domain deep learning for robust ADS-B spoofing detection via reversible mormalization and spectral patch attention

Automatic Dependent Surveillance-Broadcast (ADS-B) technology serves as a critical infrastructure in modern civil aviation, significantly enhancing airspace capacity, flight safety, and operational efficiency through high-precision real-time aircraft surveillance. However, its inherent vulnerabilities-including open signal protocols and lack of encryption—render it highly susceptible to sophisticated spoofing attacks (e.g., ghost aircraft injection and trajectory manipulation), which compromise aviation safety by generating false targets or impersonating legitimate aircraft. To address these limitations, this study proposes RDPA (RevIN Discrete-Cosine Patch Attention), a dual-domain deep learning framework that synergistically integrates temporal stabilization and spectral feature enhancement. The framework pioneers Reversible Instance Normalization (RevIN) to dynamically mitigate non-stationary distribution shifts across flight phases while preserving signal fidelity through invertible transformations. Complementarily, the Discrete Cosine Transform–based Patch Attention (DCT-PA) module decomposes trajectories into fine-grained spatio-temporal patches to localize spectral distortions induced by spoofing. Evaluations based on real-world ADS-B surveillance data demonstrate that RDPA consistently outperforms the baseline methods when using a 1% anomaly rate as the primary evaluation criterion. Particularly under this framework, the F1 score reached 99.02%. These results highlight RDPA’s robustness in detecting diverse and low-visibility spoofing behaviors and underscore its potential for deployment in operational aviation security systems.