LSTM-Transformer Fusion Variational Autoencoder (LTF-VAE) for ADS-B Anomaly Detection

Automatic Dependent Surveillance-Broadcast (ADS-B) is a critical component of next-generation air traffic management systems. However, its vulnerability to spoofing attacks stems from the broadcast of plaintext messages lacking encryption and authentication, potentially leading to severe consequences. To address this, this paper proposes an LSTM-Transformer Fusion Variational Autoencoder (LTF-VAE) for ADS-B anomaly detection. The key innovations are: (1) The encoder employs a Bidirectional LSTM for local temporal modeling combined with a 3-layer 8-head Transformer for global spatio-temporal modeling. (2) Variational inference is introduced to generate latent space distributions, constraining the model to probabilistically represent normal flight patterns. (3) The decoder utilizes a cascade structure of a single-layer LSTM and a 2-layer Transformer, reconstructing multidimensional flight parameters synchronously via fully connected layers. Experiments demonstrate that the model effectively detects diverse ADS-B anomalies under various attack scenarios, outperforming baseline methods.