A Novel Intrusion Detection Method Based on Supplement Gate Recurrent Unit for IoT

Abstract

With the rapid development of information technology, the internet of things (IoT) technology has been integrated into most people’s daily life and work. However, the IoT must confront many new security challenges. Specifically, the increase in the variety of IoT-connected devices has diversified the network. Meanwhile, the high data rates and spectral efficiency offered by 5G cellular networks facilitates the increasing capacity of IoT network traffic. Therefore, network traffic data are characterized by an expanded large scale, wide diversity, and high dimensions, which greatly affects the functionality and efficiency of intrusion detection methods. Although the existing neural network-based intrusion detection methods partially resolve the above problems, they need to execute a lot of nonlinear transformations when learning and characterizing data, resulting in a large loss of feature information. To address this problem, in this paper, we first design a new neural network model based on the gate recurrent unit (GRU), namely, the supplement gate recurrent unit (SGRU). Compared with a traditional GRU, through loss compensation, a SGRU can reduce the loss of feature information caused by nonlinear transformations when learning and characterizing network traffic data. Then, we adopt the SGRU to propose a novel intrusion detection method to monitor the security of the network. Finally, we developed the corresponding prototype system and verified its performance. The experimental results demonstrate that our proposed intrusion detection method is more accurate than previous intrusion detection methods.

A New DSGRU-Based Intrusion Detection Method for the Internet of Things

“…We can obviously see that, compared with the other four previous intrusion detection schemes, our proposed scheme performs best in accuracy, precision, recall, and F1_score. For example, the accuracy of our proposed scheme reaches 96.08%, which is 1.34%, 1.86%, 1.74%, and 1.70% higher than those of the SGRU-based intrusion detection scheme [40], GRU-based intrusion detection scheme [41], BiLSTM-based intrusion detection scheme [37], and DAE-BPNN-based intrusion detection scheme [39], respectively. For another example, the precision of our DSGRU-based intrusion detection scheme reaches 95.52%.…”

“…In this subsection, we simulate our proposed scheme and compare the performance with some previous intrusion detection solutions. Specifically, we will compare the performances among our proposed DSGRU-based intrusion detection scheme, SGRU-based intrusion detection scheme [40], GRU-based intrusion detection scheme [41], BiLSTM-based intrusion detection scheme [37], and DAE-BPNN-based intrusion detection scheme [39].…”

“…Although the above solutions can improve the accuracy of intrusion detection, they need to perform a large number of nonlinear transformations when dealing with largescale, heterogeneous, high-dimensional, and temporally correlated IoT network traffic data, thus affecting efficiency and real-time performance. Meanwhile, they suffer from the problem of local data feature loss, which can reduce accuracy and generalization ability [40].…”

“…In the following part, we will analyze the theoretical computational complexity of the DSGRU-based intrusion detection scheme, SGRU-based intrusion detection scheme [40], GRU-based intrusion detection scheme [41], BiLSTM-based intrusion detection scheme [37], and DAE-BPNN-based intrusion detection scheme [39], as demonstrated in Table 1. For simplicity, symbol m is used to represent the dimensionality of the input data and symbol n is utilized to represent the dimensionality of the hidden layer.…”

“…To evaluate the model fitting degree, we compare the loss values of the DSGRU-based intrusion detection scheme, SGRU-based intrusion detection scheme [40], GRU-based intrusion detection scheme [41], BiLSTM-based intrusion detection scheme [37], and DAE-BPNN-based intrusion detection scheme [39]. Subsequently, the comparison results are demonstrated in Table 2 and Figure 6.…”