基于Transformer-VAE的ADS-B异常检测方法

ADS-B anomaly detection method based on Transformer-VAE

针对现有的异常检测算法无法同时捕捉广播式自动相关监视(automatic dependent surveillance-broadcast,ADS-B)报文序列中的长程依赖关系、随机性以及整体性特征的问题,提出了基于Transformer变分自动编码器(variational autoencoder,VAE)的ADS-B异常检测算法,以Transformer编码器和解码器作为VAE的推断网络和生成网络来学习ADS-B报文序列的时序分布。由Transformer编码器的自注意力机制建模序列中的长程依赖关系,生成反映报文序列整体性特征的随机变量,并采用一个特殊的融合模块融合该随机变量与Transformer解码器的输出以实现报文的重构,最终通过重构误差来检测异常。实验表明,所提出的方法在不同攻击场景下,均优于相关基线算法。

Aiming at the problem that the existing anomaly detection algorithm cannot simultaneously capture the long-range dependency,randomness and overall characteristics in the automatic dependent surveillance-broadcast(ADS-B)message sequence,an anomaly detection algorithm for ADS-B based on variational autoencoder(VAE)and Transformer is proposed.The Transformer encoder and decoder are used as the inference network and generation network of VAE to learn the timing distribution of ADS-B packet sequence.The long-range dependence relationship in ADS-B packet sequence is modeled by the self-attention mechanism of Transformer encoder,and random variables reflecting the global information of ADS-B packet sequence are generated.A special fusion module is used to fuse the global random variables with the output of Transformer decoder to realize the reconstruction of ADS-B packets.Finally,anomaly is detected through reconstruction error.Experiments show that the method outperforms the relevant baseline algorithms under different attack scenarios.