A Hybrid Neural Network for Simultaneous Multi-Attack Detection in Sensor Networks

As a critical component in Internet-of-Things and Artificial Intelligence applications, wireless sensor networks suffer from various cybersecurity attacks, such as jamming and relay attacks. Since heterogeneous attacks could co-exist in practical application scenarios, existing works that only mitigate one type of attacks will expose some vulnerability space for adversaries to explore. Moreover, due to the emergence of new attacks, it is imperative to empower sensor networks with intelligent defense methods to address attack variants. In this work, we propose a hybrid convolutional-recurrent neural network (CRNN) to simultaneously detect multiple attacks in wireless sensor networks. The proposed CRNN architecture is capable of handling multi-label classification with 44% fewer neurons than RNN. Our simulation results indicate that CRNNs achieves 12% higher sensitivity in simultaneously identifying jamming and replay attacks compared to CNN. When attempting to detect multi-class attacks, CRNNs reduce the misclassification rate by 17% over CNN. Furthermore, the proposed CRNN obtains the highest temporal correlation in model learning compared to CNN and RNN.