A Transformer Based Malicious Traffic Detection Method in Android Mobile Networks

People nowadays greatly enjoy benefits and convenience from massive deployment of the mobile network in their lives. But such a large-scale internet distribution has also triggered various network threats, such as malicious information, which seriously violate people’s privacy. There are several endeavors for designing detection systems against these threats, but previous works were constrained in the two-classification scenario and biased in similar structures. Furthermore, existing works mainly focused on detection to traffic data without encryption, neglecting a larger proportion of encrypted traffic data. Therefore, devising an effective and innovative detection method is necessary to protect people’s cyber security and civil rights. In our study, we designed a detection system by introducing a novel Transformer neural network that can make classification in encrypted traffic data with higher accuracy. After implementing several experiments, our model can reach an average accuracy 95.58% with an AUC score 0.9572 in the two-classification scenario, and an average accuracy 95.54% with a maximum accuracy 96.18% in the five-classification scenario, which are better than performances from CNN and LSTM based detection methods. Hence, we concluded that our detection system based on Transformer outperforms CNN and LSTM based detectors, which possesses higher level of accuracy and more robustness in both two- and five-classification scenarios.