A Novel Deep Hierarchical Machine Learning Approach for Identification of Known and Unknown Multiple Security Attacks in a D2D Communications Network

Abstract

Intrusion Detection Systems (IDSs) have played a crucial role in identifying cyber threats for a very long time. Still, their significance has increased significantly with the advent of 5G/6G technologies, particularly Device-to-Device (D2D) communication. Multiple cyberattacks, such as Man in the Middle (MITM) attacks, Structured Query Language (SQL) injection attacks, Dictionary attacks, Distributed Denial of Service (DDoS) attacks, and others by using specific attack tools such as HULK, RUDY, and GoldenEye, that can cause rapid battery drain, rendering D2D network devices more prone to hardware failure or even to the dissolution of the D2D communication network affecting the operation and the performance of the mobile network. Using a Deep Hierarchical Machine Learning Model/Deep Hierarchical Neural Network (DHMLM/DHNN) technique, we develop an Intrusion Detection System (IDS) for D2D communication that, due to its hierarchical structure, is distinct from other comparable approaches. (i.e., Recurrent Neural Networks (RNN), Deep Neural Networks (DNN), Long short-term memory (LSTM)), has several advantages, including i) reduced training time (training time can be reduced by 56%.); ii) the ability to identify multiple types of attacks; iii) the ability to identify Zero-day/Unknown attacks (i.e., attacks that it has not seen before); iv) a more straightforward model design due to the low number of connections and neurons compared to other approaches (excluding RNN and LSTM), and; v) overall outstanding performance in terms of accuracy (i.e., 99.07%). The custom/unified data set used to train and evaluate the model was partially manually emulated and partially sampled from a large set (>95%) from the commonly used CIC-DDoS-2019 data set. The after-comparison final proposed model's 99.07% accuracy on this unified data set demonstrates the efficacy of our method. The model was also tested and demonstrated an astounding 99.63% accuracy for zero-day/unknown attacks.