An optimised Darknet traffic detection system using modified locally connected CNN - BiLSTM network
Abdullah Abdul Sattar Shaikh, M S Bhargavi, C. Pavan Kumar · International Journal of Ad Hoc and Ubiquitous Computing · 2023
The contents of the darkweb have always been a major breach of security and privacy. Due to its anonymous nature, detection of traffic from Darknet becomes difficult. A robust classifier system that accurately predicts and classifies such traffic is a necessity. This research work aims to study the effects of the convolutional-long-short-term memory (CNN-LSTM) system of classification of Darknet through various deep layer modifications on the Nadam optimiser. Experimentations were carried out on different combinations of locally-connected CNNs (LcCNN) and bi-directional LSTM (BiLSTM) to improve accuracy. Data was subjected to various levels of synthetic minority oversampling techniques (SMOTE) to reduce overfitting, data imbalance and achieve better generalisation. A custom decaying call-back function implemented, cut down the learning rate by half and tended to improve accuracy. Results obtained outperformed the base CNN-LSTM system for traffic categorisation with an improved accuracy of 92.57% from 89% using the custom LcCNN-BiLSTM architecture.