Lightweight anomaly detection model for UAV networks based on memory-enhanced autoencoders

Abstract

In order to solve the problems of high energy consumption and high reliance on manual annotation data of traditional intelligent attack detection methods in UAV networks, a lightweight UAV network online anomaly detection model based on a double-layer memory-enhanced autoencoder integrated architecture was proposed. The message queue based on the operating system was used for data packet caching to achieve persistent processing of high-speed data streams, which effectively improved the stability and reliability of the model. The composite statistical characteristics of the data flow were calculated based on the damped window model, and the memory complexity in the calculation process was reduced in an incremental update manner. The hierarchical clustering algorithm was used to divide the composite statistical features, and the separated features were input to multiple small memory-enhanced autoencoders in the integrated architecture for independent training, which reduced the computational complexity and solved the problem of false negatives caused by the overfitting of the reconstruction effect of the traditional autoencoder. Experiments on public data sets and NS-3 simulation data sets show that while ensuring lightweight, the proposed model reduces the false negative rate by an average of 35.9% and 48% compared with the baseline method.

Keywords

• UAV network;anomaly detection;lightweight online detection;memory-augmented autoencoder