DEEP LEARNING-BASED INTRUSION DETECTION FOR REAL-TIME IDENTIFICATION OF NETWORK SECURITY THREATS

Abstract

Rapid development in computer network interconnections has resulted in more sophisticated and frequent attacks on the networks. Classical intrusion detection systems face challenges when dealing with advanced and novel attacks that are not seen before. Deep Learning offers a potential way of automatically learning the complicated attack patterns from network traffic data. This research work seeks to design a deep learning framework for detecting malicious network traffic. A data set consisting of 52,480 instances of network traffic data under ten different attack classes was used. The data set was preprocessed using data normalization, feature selection, and removal of duplicates. CNN, LSTM, Random Forest and SVM models were created and analyzed. Model efficiency was evaluated based on accuracy, precision, recall, F1 score, false-positive rate, and detection time. The overall accuracy, precision, recall, and F1-score of the suggested CNN-LSTM architecture was 97.3%, 96.8%, 97.1%, and 96.9%, respectively. The false positive ratio of the proposed method was reduced to 2.4% as compared to the range of 4.8% to 7.1% achieved through conventional machine learning techniques. The total number of correctly classified network attacks was 51,064 out of 52,480 instances. Threat detection latency was found to be around 38 ms; this is 41% less than that in the baseline LSTM-based model. The highest detection rate was obtained for denial of service attacks (98.6%) and malware attack (97.9%).The proposed deep learning framework has shown highly accurate detection results and low computational latency for network intrusion detection.