PaBiL-Net: Parallel Bidirectional LSTM Network for Mobile Encrypted Traffic Classification

With the rapid popularity of smart handheld devices, the demand for the classification of mobile encrypted traffic is growing quickly. The data feature selection strategy is an important factor affecting classification performance. However, most classification models in recent years have used a fixed feature selection strategy, making it difficult for the models to adapt to new datasets and application scenarios. We propose an end-to-end classification model called parallel bidirectional LSTM network (PaBiL-Net) to address this issue. The model arranges similar neural networks in parallel and processes different features separately in each parallel path, thus it is more flexible in feature selection and can easily adjust the type and number of features computed by the model. We compare PaBiL-Net with 3 baseline methods by conducting mobile encrypted traffic classification experiments on 2 publicly available datasets. The results of 5-fold cross-validation show that PaBiL-Net outperforms the baseline models with an increase in accuracy of 4.0939% and 2.2061% on the 2 datasets respectively. Additionally, the macro F1 increased by 3.9029% and 2.2933%, respectively.