Integrated Framework for Intrusion Detection Through Adversarial Sampling and Enhanced Deep Correlated Hierarchical Network

The internet provides everyone with access to data, information, and knowledge that aids in our professional, interpersonal, and economic growth. The internet serves various purposes, but how we utilise it in the daily lives varies greatly according to our specific needs and ambitions. The exponential growth of the internet has led to an increase in digital communication among the general public. The number of intrusion occurrences has increased significantly. Internet security is a vital aspect of an organization's effectiveness because cyber-attacks may affect corporate operations. Hence, there is essential for the development of a security mechanism to handle these systems [1, 2]. According to a recent analysis, the frequency of network attacks has grown substantially, and researchers' interest in network intrusion analysis has risen as a result. In the network security domain, the detection of intrusions has evolved into a significant research problem [3, 4].

The technique of detecting anomalous abnormal actions on computer systems is known as intrusion detection. The main aim of an Intrusion Detection System (IDS) is to identify users' behavior as normal or abnormal based on the data they communicate. Firewalls, data encryption, and authentication techniques were all employed in traditional security systems [5, 6]. Anomaly detection and misuse detection are the most widely used in intrusion detection systems today, although these two approaches have the limitations of minimum detection rates and large false-positive rates. Detection approaches based on artificial intelligence has a crucial in the development of IDS [7, 8]. Researchers employ machine learning to identify several forms of attacks to construct an effective intrusion detection system.

A hybrid model with feature analysis technique [9] was introduced for IDS. A combination feature selection and Support Vector Machine (SVM) classifier ID model [10] was built using Chi-square, Modified Naive Bayes (MNB), SVM, and LPBoost. A majority votings of LP Boost MNB and SVM was used to forecast the intrusion label effectively. A combined classifier model based on tree-based methods [11] was proposed. Naive Bayes Tree, mRandom Tree and Decision Tree were combined. As a result, integrating classifiers using the sum rule scheme can produce better results than using individual classifiers. SVM, RF and Extreme Learning Machine (ELM) were developed [12] for intrusion detection and finally proved ELM outperform then other two classifiers.

Classical Machine Learning (ML) techniques are based on narrow learning, but as network data continues to grow, a vast volume of non-linear network data presents new obstacles for intrusion detection [13]. This paper analize ML based methods and Deep Learning (DL) based approaches and provide a new solution for IDS system.

The remaining sections of this article include: Section 2 studies the recent work related to the intrusion detection categorization methods. Section 3 explains the proposed methodology of proposed classification frameworks and Section 4 displays their efficiencies. Section 5 concludes the entire work and suggests the future scope.

This paper built a deep hierarchical network that integrates the proposed Cross-correlated Convolution Neural Network (CCNN) and Bi-directional LSTM (BiLSTM), named as CCNN-BiLSTM for learning the spatial and temporal aspects of NTD to deal with the difficulty of data features. Because of its strong performance in automatic feature extraction, CCNN has steadily been applied to network identification. Considering that sample attacks are time sequence data, which is intra-dependent. Automatically measure the time sequence features, this research work used BiLSTM for learning. The main contribution of this research work is given below.

(1) To overcome the data imbalance problem, the research developed an Adversarial Sampling strategy. Enhanced Generative Adversarial Networks EGAN is proposed in this research as a way to increase minority sample sizes. In this method created a balanced dataset for model training. Hence, the time for training the model is cut in half.

(2) The deep hierarchical network model, which combines Cross-correlated Convolution Neural Network (CCNN) and BiLSTM, is proposed. This approach extracts the data's features with reliability. The CCNN obtained a model performs better after training.

1.png

Figure 1. System architecture for proposed NID model

Figure 1 depicts the system architecture of the proposed Network Intrusion Detection (NID) model. To address the imbalance data and the difficulty of the features, the original data is initially applied to Adversarial Sampling process to provide balanced data. Finally, classification is performed in a Deep Hierarchical Network (DHN). It is divided into four sections, which are Adversarial Sampling, model training, and classification.

The performance of the classification model is influenced by the imbalanced data. As a result, this research work employs the EGAN algorithm to raise the minority samples.

The problem of low detection rate and difficulty of feature learning in a small number of attack categories (U2R; R2L) are not considered in existing intrusion detection. Hence to handle the issue of feature learning in a small number of attack categories, a generative adversarial network method is proposed to identify more features.

The NID dataset instances contain too many 0 elements that are not helpful for network learning and can speed up the convergence of the network. This is not resolved by the data normalization approaches. Hence to prevent the over fitting problem due to 0 elements, a cross-correlation-feature layer is included in the DNN model. The proposed CCNN and BiLSTM are help to retrieve the data’s temporal and spatial properties to increase the classification accuracy.

Resistance training, created a model with better classification performance, and the model is utilized to test set categorization, yielding better classification outcomes.

3.1 Balanced data set construction using adversarial sampling

A typical imbalanced data classification challenge is NTD, which consists of a minimum number of anomalous traffic and significant volume of regular traffic. Although the prediction is accurate in this situation, the overall error is minimized. This work proposed an Enhanced Generative Adversarial Networks (EGAN), which is enhanced by incorporating an inference network into the conventional GAN model. It allows the classification methods to consider the impact inputs from both data and latent space. It learns the latent representation by combining autoencoder structure with a normal GAN framework. The generative net (G) and the discriminative net (D) are the two kinds of network in GAN. In the training stage generative net (G), the G attempts to generate fake data based on input priority during the training stage. The traditional GAN has some significant issues, which are:

Model parameters fluctuate and destabilize and never coverage;

The generator fails, resulting in a restricted number of sample variations;

The discriminator becomes too successful, and the generating gradient vanishes, leaving the learner with no information;

Over fitting is caused by an imbalance between the discriminator and the generator.

Hyper - parameter choices are quite delicate. To overcome these issues, this research work combined Basic Iteration Method (BIM), Fast Gradient Sign Method (FGSM), and Projected Gradient Descent (PGD) algorithms. Initially, the FGSM method is to create an adversarial sample from NSL-KDD Dataset. This technique performed an updating of 1-step gradient in the direction of the gradient for each dataset input. The BIM is the second method [26], which performs a greater optimization of the FGSM over multiple rounds with modest alterations [27]. Any feature of the input parameters is cropped in any cycle to avoid too significant a change on each feature. The PGD is a version of the FGSM attack that does not include the FGSM's random start characteristic [28]. For identifying such adversarial examples, PGD is an effective first-order technique, which also used to address optimization issues involving the simultaneous minimization of several objective functions. To generate the adversarial samples PGD can produce more accurate perturbation by using a uniform random perturbation as an initialization. Then it runs in the multiple BIM iterations to identify the adversarial samples. All these three approaches rely on the model gradient and are model dependent. Procedure was repeated with the BIM and PGD adversarial samples for getting higher accuracy.

2.png

Figure 2. Adversarial sampling by EGAN using EGSM, BIM and PGD process

The proposed EGAN is used in the Adversarial Sampling process, which aims to raise the minority samples in the imbalanced NTD. Figure 2 represents the Adversarial Sampling Process. EGAN is used of three neural networks: a generator (g), an encoder (i), and a discriminator (d), with g and i forming the autoencoder. The encoder i compresses real data samples (a) into a latent representation (b), while the decoder g reconstructs the encoded data back into the original data (a). By learning the joint posterior distribution p(a|b), this autoencoder structure can reconstitute the original data, improving the model's stability by decreasing mode collapse caused by GAN's inability to infer the mapping of actual data to latent data. The auto encoder may also assist with operations at the level of abstraction by utilizing the encoder to learn the inference (a b), where a (a) represents actual data spaces and (b) represents latent spaces, of the latent representation of high-dimensional data spaces. In terms of the advantages of learning latent representations in Eq. (1):

$\max _{g, i} \min _g e(d, i, g)=i_{a \sim p a}\left[i_{a \sim p i}(. \mid a)[\log d(a, b)]\right]+i_{b \sim p b}\left[i_{a \sim p g}(. \mid b)[1-\log d(a, b)]\right]$

$=i_{a \sim p a}[\log d(a, i(a))]+i_{b \sim p b}[1-\log d(g(b), b)]$     (1)

where, $p a$ and $p b$ are the sharing over the data. The $p i(. \mid a), p a(. \mid b)$ and $p g(. \mid a)$ and $p b(. \mid b)$ are the joint distributions of $p i, p a, p g$, and $p b$, which are demonstrated using encoder and generator.

Algorithm 1 Adversarial Sampling Algorithm

3.2 CCNN for extraction of spatial features

The proposed Cross-correlated Convolution Neural Network (CCNN)can automatically extract the features of objects better than traditional methods. The features delivered from EGAN is given to the CCNN for training. Because the characteristics have spatial locality at dissimilar places, pooling layers must be used to assemble the information of the characteristics at different positions to some extent to eliminate the dimension and overfitting should be avoided. As a result, CCNN is well suited to extracting geographical information from large networks' NTD.

3.png

Figure 3. CCNN network structure model

Figure 3 depicts the CCNN network structure model. The procedure for mining the spatial characteristics using CCNN is given below The CCNN the introduction of the idea of local perception, the convolution kernel can be shared by all neurons, and the collection of weights is determined by the number of convolution kernels. As a result, the collection of weights is drastically eliminated, while computational efficiency is greatly improved. The convolution function is denoted by the symbol in Eq. (2):

$h_t=C\left(h_{t-1} \otimes h_t+b_a\right)$     (2)

where, h_a is the character map of layer, bias layer is $a,(a D 1 ; 2 ;:: n) b_a$, and $\otimes$ is the convolution function and c (x) is the activation function. This research work used ReLU as CCNN activation function. After the convolution, the pooling layer combines the features in the local neighborhood to generate new features. It is able to decrease the feature map h_t size and preventing over-fitting, which is defined in Eq. (3) as follows:

$h_t=\operatorname{pool}\left(h_{t-1}\right)$     (3)

where, h_t must be transformed into a vector after numerous convolution and pooling layers. The completely linked layer can then be used to produce output u_a. As a result, we can extract the geographical properties of the NTD using CCNN. While CCNN can extract spatial characteristics effectively, it struggles to learn sequence correlation information and cannot overcome the problem of long-term information dependence. As a result, network intrusion detection accuracy using solely CCNN has to be increased.

3.3 Extraction of temporal features by BiLSTM

The LSTM contains a memory module which can decide whether to store information in memory and when to forget it LSTM can effectively extract time series with significantly larger intervals and delays, effectively resolving gradient disappearance and training challenges. However, the importance of information after characteristics is not completely taken into consideration because the LSTM could only accept the sequence data in one way. As a result, BiLSTM is utilized instead of LSTM to present the upcoming information. Figure 4 demonstrates Fundamental Structure of BiLSTM.

4.png

Figure 4. Fundamental structure of BiLSTM made up using stacking of 2 LSTM

Figure 4 demonstrates Fundamental Structure of BiLSTM. The hidden loop layer, input layer, and an output layer comprise a standard LSTM network framework. The cyclic hidden layer, unlike the standard recurrent neural network, is mostly made up of nodes that represent neurons [23]. The memory module is the fundamental element of the cyclic hidden layer LSTM. The input gate, forget gate, and output gate are among the 3 adaptive multiplication gating units of this memory module. Each neuron node in the LSTM performs the following mathematical operation:

The input gate is activated at time t according to the previous cell's output result h_a in represent in Eq. (4). The current input x_t decides whether or not to use calculation to update the current information in the cell. The weight is B, while the bias of neurons is b:

$h_t=\operatorname{sigmoid}\left(B_i \cdot\left[h_{t-1}, x_t\right]+b_i\right)$     (4)

The information is retained or discarded using a forget gate, and the last instantaneous hidden layer output t_i-1 and the input present time in Eq. (5):

$C_i=\operatorname{sigmoid}\left(B_c \cdot\left[h_{t-1}, x_t\right]+b_c\right)$     (5)

The present value of a potential memory cell is determined by the previous moment's output result t_i-1 and input data of the LSTM hidden layer cell. The present value of a potential cell S_i and its state S_i-1, as well as the forget gate and input gate are currently altering the memory cell state value S_i. The unique feature is the element-by-element matrix multiplication present in Eq. (6) and Eq. (7):

$y_i=\tanh \left(B_S \cdot\left[h_{t-1}, x_t\right]+b_S\right)$     (6)

$y_i=C_i+S_{i-1}+h_t * S$     (7)

Calculate the output gate o_t, which is utilized to regulate thevalue of the cell status. The last cell's output is h_t that can be written as Eq. (8) and Eq. (9):

$o_t=\operatorname{sigmoid}\left(W_0 \cdot\left[h_{t-1}, x_t\right]+b_O\right.)$     (8)

$h_t=o_t \tanh \left(S_t\right)$      (9)

The BiLSTM is made up of 2 LSTM networks, one backwardand the otherforward. The forward LSTM hidden layer is in charge of extracting forward features, while the backward one is in charge of extracting backfeatures in Eq. (10), Eq. (11) and Eq. (12):

$h_t^{\text {backward }}=L S T M^{\text {backward }}\left[h_{t-1}, x_t\right]+S_{t-1}$     (10)

$h_t^{\text {forward }}=L S T M^{\text {forward }}\left[h_{t-1}, x_t\right]+S_{t-1}$     (11)

$H_t=\left[h_t^{\text {backward }}, h_t^{\text {forward }}\right]$     (12)

Prior and after it is applied, the BiLSTM model can effectively understand the importance of every attribute in the sequence information. As a result, more detailed feature information is gathered. BiLSTM's state at time t comprises both forward and backward output [22].

3.4 Enhanced deep correlated hierarchical model

An innovative CCNN was proposed in this research work to improve assessment metrics of imbalanced abnormal traffic data classification in intrusion detection. Finally, we offer some modified variants of the CCCN that may balance detection efficiency and time efficiency to address the algorithm's time efficiency. In the experimental part, we concentrate on enhancing each category of imbalanced anomalous flow data evaluation metrics rather than the total evaluation metrics. By merging CCNN and BiLSTM networks in hierarchical network architecture, the temporal and spatial elements of traffic data may be obtained concurrently. Due to the differences in the CCNN and BiLSTM inputs, the derived spatial characteristics are altered at the CCNN output to fit the BiLSTM network's input. In the CCNN based BiLSTM method has some issues, which are CCNN lacks the capacity to be spatially independent to the input data and does not encode the object's location and orientation.

Both CCNN and BiLSTM are deep learning techniques that are representational. The Figure 5 gives DHN. In the spatial dimension, CCNN may retrieve data characteristics. BiLSTM has the property of retaining relevant historical data for an extended period and achieving data feature extraction at the temporal level. It is vital to examine the feature connection at the spatial level while extracting features for a network intrusion detection system, in addition to measuring the law of change at the temporal level. As a result, this article used CCNN and BiLSTM to retrieve the features before building a DHN. All convolution kernels in the CCNN are 3*3 in size because a lesser convolution kernel can efficiently lower the model's difficulty Furthermore, the larger convolution kernel's comparable responsive field can be obtained by stacking multilayer 3*3 convolution kernels.

5.png

Figure 5. Enhanced deep correlated hierarchical structure model

Assuming that X_a-1 is the current layer's input feature map and that the convolution kernel's input size is 3*3, the current layer's output feature map X_a-1 is stated as Eq. (13):

$Y_a=\omega_a * X_{a-1}+b_a$      (13)

The letter b_a stands for the bias word. Due to the huge discrepancy in the distribution of flow data samples from different categories, the network adapts effectively to anomalous categories with large sample sizes for highly imbalanced abnormal own data, but the detection impact is often poor for own with few samples. Batch normalization (BN) was used to prevent overfitting [20]. BN reduces the internal covariate shift of the input data, which speeds up the convergence of the training deep network model. BN decreases the correlation of each feature dimension when dealing with unbalanced data, allowing the network to quickly comprehend the categories despite tiny sample sizes. By computing the mean and variance of the samples in an input mini-batch, the data for that mini-batch is normalized. The following is the BN transformation procedure in Eq. (14) and Eq. (15):

$\mu_D=\frac{1}{N} \sum_{a=1}^n y_a$     (14)

$\sigma_D^2=\frac{1}{n} \sum_{a=1}^n\left(y_a+\mu_D\right)^2$     (15)

The mean $\mu_D$ and variance $\sigma_D^2$ of the mini-batch are calculated for each of the $n$ input examples, and the variance and mean are then normalized. The unique procedure is as follows in Eq. (16):

$\widehat{y_a}=\frac{y_a-\mu_D}{\sqrt{\sigma_D^2+\varepsilon}}$      (16)

where, $\mu_D$ represents the mean of all examples in a less-feature batch's dimensions, and $\sigma_D^2$ represents the variance of whole examples in the less-batch. To obtain $\widehat{y_a}$, the example $y_a$ is normalised by $\mu_D$ and $\sigma_D^2$. The normalize procedure is used in every input layer to ensure that all input data has the same distribution regardless of sample variability. The features are scaled and moved to increase the generalization of new data by introducing two learnable parameters and requiring the model to learn the original sample distribution to improve the resilience of network learning present in Eq. (17):

$D N_{\gamma, \beta}\left(y_a\right)=\gamma^{y_a}+\beta$     (17)

The nonlinearity is added using an activation function after the data has been batch normalized. All activation functions in our network structure use the ReLU function to speed up network convergence and resolve the impacts of gradient is-appearance and gradient explosion. As a result, the last output feature map is written as Eq. (18):

$z_a=g\left(D N_{\gamma, \beta}\left(\omega_a * Y_{a-1}+b_a\right)\right)$     (18)

The activation function is denoted by the letter g. Finally, we utilize the Softmax function for our classification, and each category's sample output is stated as Eq. (19):

$x_a=\frac{c^{z_a}}{\sum_{d=1}^e c^{2^d}}$     (19)

Algorithm 2 Deep Correlated Hierarchical Model