A neural network model for attacker detection using GRU and modified kernel of SVM

Abstract

Over past few decades neural network changed the way of traditional computing many different models has proposed depending upon data intensity, predictions, and recognition and so on. Among which Gated Recurrent Unit (GRU) is created for variety of long short-term memory (LSTM) unit, which is part of recurrent neural network (RNN). These models proved to be dominant for range of machine learning job such as predictions, speech recognition, sentiment analysis and natural language processing. In this proposed model, a support vector machine (SVM) with modified kernel as final output layer for prediction is used instead of traditional approach of softmax and log loss function is used to calculate the loss. Proposed technique is applied for binary classification for intrusion detection using honeypot dataset (2013) network traffic sequence of Kyoto University. Results shows a prominent change in training efficiency of ≈89.45% and testing efficiency of ≈88.15% when compared with softmax output layer. We can conclude that linear SVM with modified kernel as output layer outperform compared with softmax in prediction time.