As the Internet applications are growing rapidly, the intrusion detection system is widely used to detect network intrusion effectively. Aiming at the high-dimensional characteristics of data in the intrusion detection system, but the traditional frequent-pattern-based outlier mining algorithm has the problems of difficulty in obtaining complete frequent patterns and high time complexity, the outlier set is further analysed to get the attack pattern of intrusion detection. The NSL-KDD dataset and UNSW-NB15 dataset are used for evaluating the proposed approach by conducting some experiments. The experiment results show that the method has good performance in detection rate, false alarm rate, and recall rate and effectively reduces the time complexity.

1. Introduction

1.1. Intrusion Detection System

With the rapid development of modern information technology, network security has become the focus of attention. How to effectively detect the types of intrusion attacks, as well as the security of the early warning and protection system, has become one of the research directions of network security. Intrusion detection systems (IDSs) are most widely used in the world for identifying and detecting the intruders in computer networks, Internet, and cloud networks. The intrusion detection system analyses the network data collected by the computer system and the key points in the network, so as to find out the behaviour of violating the security policy and the traces of attacks and monitor and detect the network intruders. The IDS can be used to detect different types of attacks on the network, but the traditional firewall cannot perform these attacks well.

Generally, the intrusion detection system can be roughly divided into two categories according to its detection methods, namely, an anomaly detection system and detection system. Anomaly detection is also known as behaviour-based system detection, which detects the abnormal behaviour of the system to discover intrusion behaviour. Misuse detection is a knowledge-based detection or feature-based detection technology, whose premise is that intrusion behaviour and normal network access have different data characteristics. The intrusion detection system is divided into two stages, namely, the preprocessing stage and intrusion detection stage. By developing the intrusion detection system, the intrusion behaviour can be identified effectively.

1.2. Outlier Detection

Outlier mining is an important research direction in the field of data mining. Outlier data do not conform to the general rules of data and are not consistent with other parts of the data. It is those small-scale objects that are far away from other objects in the dataset. Although outlier data are “abnormal data” which are inconsistent with normal data, outlier detection can provide important information in some applications.

There are many reasons for outliers. Generally speaking, they can be divided into two situations: first, they are indeed caused by human or detection equipment errors; second, they are caused by the nature of things themselves, and they are the data reflection of the real nature of things. The outlier analysed in this paper belongs to the second case. The outlier data generated by human operation are significantly different from the normal network behaviour, in order to find the real potential valuable knowledge through outlier mining.

In the real network activities, most of the network behaviours are normal, the intrusion behaviour can be regarded as the abnormal phenomenon of the amount of data far less than the normal behaviour, and the data corresponding to the normal behaviour and the intrusion behaviour have different data characteristics. Based on the characteristics of intrusion behaviour data, intrusion behaviour can be regarded as “outlier” data [1].

1.3. Association Rule Mining

Association rule mining, as an important part of data mining, has been a hot research topic. Association rules are a collection of items in the database that exceed the specified minimum support and minimum confidence. Association rules are usually expressed as X ⇒ Y, support = s, and confidence = c, in which X is the precondition of the rule, Y is the conclusion of the rule, the support s represents the frequency of the rule, and the confidence c represents the strength of the rule.

The goal of association rule mining is to find out all the strong association rules. The mining process is divided into two steps:Step 1: all rules that are not less than the minimum support threshold s are found, i.e., all frequent patternsStep 2: by setting the confidence threshold c, the conversion rule is used to filter out the set of items less than the minimum confidence c, and the corresponding association rules are obtained

In this paper, it is only needed to get the maximum frequent patterns based on frequent pattern, so it is only needed to complete Step 1 to get the frequent pattern.

1.4. Maximum Frequent Pattern

If the maximum frequent pattern needs to be explained, the concept of supersets must be introduced first, which is defined as follows: if every element in set S2 is in set S1 and set S1 may contain elements that are not in S2, then set S1 is a superset of set S2. If set S1 is a superset of set S2, then set S2 is a true subset of set S1, and vice versa.

With the superset, the maximum frequent pattern is defined as follows: if all supersets of frequent pattern X are nonfrequent patterns, then X is called as a maximum frequent pattern.

With the increasing number and dimension of collected data in the intrusion detection system, researchers have proposed a variety of typical high-dimensional outlier mining algorithms for the complexity, sparsity, and diversity of high-dimensional data. Among them, outlier mining based on frequent pattern is widely used in intrusion detection because of its easy-to-understand nature and low time complexity. On the basis of frequent-pattern-based outlier mining algorithm, using the concept of maximum frequent pattern in association rules, an improved high-dimensional outlier mining algorithm based on the maximum frequent pattern is proposed in this paper. The algorithm transforms frequent pattern mining into maximum frequent pattern mining. On the premise of good detection performance, the time complexity is reduced.

2. Literature Survey

In the real network, the data are high dimensional in the intrusion detection system. Some researchers proposed the means to reduce the dimension of high-dimensional data with the way of feature extraction or feature selection and then analysed the processed data with the traditional data mining methods.

Ganapathy [2] proposed an intelligent algorithm for feature selection and classification to design an effective intrusion detection system, which can be used to provide security to networks effectively.

Tian et al. [3] proposed a hierarchical outlier detection model based on PCA, an anomaly data model based on PCA was established based on normal data to filter data firstly, and then, the abnormal data types were analysed to detect both anomaly and misuse attack.

Zyad et al. [4] proposed a way to use the trimmed average vector to estimate the average vector on the basis of PCA, so as to make the trimmed PCA have better robustness.

To solve the problem of high-dimensional data in IDS, Riyaz and Ganapathy [5] proposed a new fuzzy rule and information gain ratio-based feature selection algorithm (FRFSA), and the existing classifiers called SVM and LSSVM were used for effective classification. The experimental result shows that the proposed work exceeds the performance measure when compared to the existing algorithms on classification for feature selection.

Nancy et al. [6] proposed a dynamic recursive feature selection algorithm for feature selection and then used an intelligent fuzzy temporal decision tree algorithm to effectively detect intruders, which can effectively reduce the false positive rate, energy consumption, and delay of the system.

The method of dimension reduction can eliminate some features and reduce the time complexity, but each feature represents a different outlier value. If the features are selected incorrectly, it will get the wrong outlier value, which will produce an approximate result that is not suitable for future calculation [7]. The complexity, sparsity, and diversity of high-dimensional data restrict the traditional mining algorithm. When dealing with high-dimensional data, data mining algorithms suitable for low-dimensional data usually encounter the problems of algorithm efficiency reduction and the traditional definition based on distance and density is invalid, which reduces the accuracy of intrusion detection [8].

Researchers have proposed intrusion detection methods for high-dimensional data. Zhang et al. [9] proposed SPOT technology for anomaly detection in a high-dimensional data network data stream, which has good detection effect.

Prajapati and Bhartiya [10] proposed a nearest neighbour search algorithm based on the advantages of K-mean algorithm and fuzzy C-mean (FCM) algorithm to solve the problem of uneven data and rigid clustering in high-dimensional data, which can realize nearest neighbour search in a shorter time.

In general, the “attack” data in intrusion behaviour are regarded as abnormal data, and outlier mining is to mine those abnormal data which deviate from normal behaviour in large-scale data, so outlier mining is very important for analysing intrusion behaviour. For high-dimensional outlier mining, researchers have proposed several typical mining algorithms: outlier mining algorithm based on spatial projection [11, 12], outlier mining algorithm based on a hypergraph model [13, 14], and outlier mining algorithm based on frequent patterns. The outlier mining algorithm based on frequent patterns is simple, easy to understand, and has lower time complexity than the previous two algorithms, so researchers have conducted extensive research.

In the early stage, He et al. [15] proposed an outlier mining algorithm based on frequent patterns (FindFPOF) and proposed a measurement factor of frequent pattern outlier factor (FPOF). It is believed that the less frequent the patterns contained in a data record, the more likely they would be an outlier, so outliers could be found by calculating the frequent pattern factor of each data.

Zhou [16] proposed a new metric called weighted frequent pattern outlier factor for categorical data streams based on FindFPOF and proposed a fast outlier detection method for high-dimensional categorical data streams based on frequent pattern (FODFP-Stream), which has good applicability and validity.

Wang and Tang [17] proposed an algorithm based on frequent patterns-NFPOF, which further accurately locates abnormal properties of each outlier data through the related attributes of frequent patterns.

Yuan et al. [18] proposed a weighted frequent-pattern-based outlier (WFP-Outlier) to solve the problem whose weights seriously affect outlier detection results, which can find implicit outliers from weighted data streams.

To solve the problem of being incapable of detecting new type of attacks, Jaisankar [19] proposed a new intelligent-agent-based IDS using Fuzzy rough-set-based outlier detection and Fuzzy rough-set-based SVM. The system adopted Fuzzy rough-based SVM in our system to classify and detect anomalies efficiently. The experimental result shows that the proposed intelligent-agent-based model improves the overall accuracy and reduces the false alarm rate.

In order to solve the problem of high false positives, Ganapathy [20] proposed a new intrusion detection model using a new Weighted-Distance-Based Outlier Detection (WDBOD) algorithm and an Enhanced Multiclass Support Vector Machine algorithm, which has low false alarm rate and high accuracy.

Combined with attribute selection, outlier detection, and the enhanced multiclass support vector machine classification method, Ganapathy et al. [21] proposed a new intelligent-agent-based intrusion detection model for mobile ad hoc networks. Using the proposed Intelligent Agent Weighted Distance Outlier Detection algorithm and Intelligent-Agent-based Enhanced Multiclass Support Vector Machine algorithm, the proposed model can detect anomalies with low false alarm rate and high accuracy.

To sum up, high-dimensional outlier mining based on frequent patterns plays a very important role in intrusion detection, but there are two problems in the algorithms based on frequent patterns. First, it needs to mine the complete frequent patterns in the dataset, but it is very difficult to find the complete set of frequent patterns in high-dimensional data. Second, the time complexity of mining algorithm for frequent patterns is exponentially related to the dimension of data, the higher the dimension, the greater the time complexity. High-dimensional outlier mining algorithm based on frequent patterns has the problems of difficulty in obtaining complete frequent patterns and high time complexity. So, a high-dimensional outlier mining algorithm based on the maximum frequent pattern factor is proposed in this paper using the concept of maximum frequent pattern factor in association rules. Also, the algorithm is applied in intrusion detection, which reduces the time complexity on the premise of ensuring good detection performance.

3. Proposed Work

3.1. Relevant Theories

We let D = {tl, t2,…,tn} be a dataset containing n network behaviour records t, and tk is called a transaction. Also, I = {il, i2,…,ip} is the collection of all attributes in the network behaviour record, and im is called an item.

Definition 1. Itemset: any subset X of I is called the itemset of D. We let tk be a transaction of D, and X is a itemset of D; if , then the itemset D is contained in the transaction tk.

Definition 2. Support: the support number of itemset X is represented as the number of transactions that contain itemset X in dataset D and is recorded as . The support of itemset X is recorded aswhere is the total number of transactions in dataset D.

Definition 3. Frequent pattern: if the support (X) is not less than the minimum support (MinSP) which is specified by the user, then X is a frequent pattern; otherwise, it is an infrequent pattern.

Theorem 1. X, Y are set as itemsets in dataset D; then,(1)If , then support (X) ≥ support(Y)(2)If and X is not a frequent pattern, then Y is not a frequent pattern(3)If and Y is a frequent pattern, then X is a frequent patternY is set as a maximum frequent pattern because , and Y must be a frequent pattern; it can be seen from Theorem 1 that X must be a frequent pattern, that is to say, all frequent patterns have been implied in the maximum frequent patterns. Therefore, the problem that the complete set of frequent patterns must be found in the outlier mining algorithm based on frequent patterns can be transformed into finding the maximum frequent patterns. It not only solves the difficulty of finding the complete frequent pattern sets but also greatly reduces the number of frequent patterns n, thus reducing the time complexity of the algorithm.

3.2. Data Discretization

The data types of attributes in a dataset can be divided into textual data and numerical data, and numerical data also can be divided into discrete data and continuous data. The data type in outlier mining based on maximum frequent patterns must be discrete data, so it is necessary that continuous attributes are converted to reliable accurate data suitable for data mining by data discretization.

The discretization of numerical attribute is to divide the continuous data into a number of finite discretization intervals. The usual discretization methods include the equal-width method, the equal-frequency method, and the method based on clustering. Clustering is an unsupervised algorithm; according to the distribution characteristics of data to determine how to divide the interval of attribute values, as far as possible to reduce manual intervention, it has been widely used in practice. After clustering, the objects in the same clustering pattern have a high similarity and are quite different from the objects that do not belong to the same clustering pattern, and data in a same clustering pattern are often treated as a whole in many practical applications. In order to minimize the intervention of human factors, the method based on clustering is adopted for data discretization in this paper.

The discretization method based on clustering has two steps:(1)Continuous attributes are clustered by the clustering algorithm(2)Patterns obtained by clustering are processed, and continuous attribute values in the same clustering pattern are uniformly marked as one value

Among them, clustering is the key step in discretization. K-means is a classical clustering algorithm based on partition, which has good effect and is widely used in practice. However, K-means algorithm is very sensitive to the number of clustering K and the selection of initial clustering centre.

For the sensitive problem of K value, the elbow method can be used to determine the optimal K value because K value is not fixed and unique in the process of discretization. The core idea of the elbow method is when K is less than the optimal number of clustering, an increase in K value will greatly increase the degree of aggregation of each clustering, so the decrease range of SSE will be very large. When K reaches the true number of clustering, the return of aggregation degree obtained by an increase in K will decrease rapidly, so the decrease degree of SSE will decrease sharply, and if K value is increased continuously, the change of SSE will tend to be gentle, that is to say, the relationship graph between SSE and K is the shape of an elbow, and the corresponding K value of this elbow is the optimal number of clusters.

The square sum of error (SSE) of the core index of the elbow method is defined aswhere Ci: the ith clustering, p: sample points in Ci, mi: the centroid of Ci (mean value of all samples in Ci), and SSE: clustering error of all samples, representing the quality of the clustering effect.

For the sensitive problem of the selection of an initial cluster centre, the maximum distance method is used to select K samples as the initial centre points based on the fact that the farthest sample points are most unlikely to be divided into the same cluster.

3.3. The Proposed Algorithm

The concept of maximum frequent pattern factor (MFPOF) is proposed based on the frequent pattern factor (FPOF) in FindFPOF algorithm.

Definition 6. Maximum frequent pattern factor (MFPOF): MFPS (D, MinSP) is the maximum frequent pattern sets in dataset D that meets a given minimum support threshold. The MFPOF of each network behaviour record t is defined aswhere is the number of the maximum frequent patterns in frequent patterns and the support(X) is the support of a maximum frequent pattern X.
The description of the high-dimensional outlier mining algorithm based on maximum frequent patterns (MFPOF-OM) is shown as Algorithm 1.

Input: D//network behaviour dataset
 MinSP//minimum support threshold
k//number of outliers threshold
Output: k network behaviour outlier data records
 // Stpe 1–3: mining the maximum frequent item sets based on PF-Tree Algorithm
 Step 1: To D, the HeaderTable (D) is generated to satisfy the MinSP;//Calculating the header table of PF-tree
 Step 2: To D, the frequent item set tree is generated to satisfy the given MinSP by using the PF-Tree Algorithm, and denoted as: T;//Obtains frequent item set tree according to the PF-Tree algorithm
 Step 3: Obtains maximum frequents item sets based on an improved PF-Tree, and obtains MFPs (D, MinSP) and support (X)//Obtains maximum frequents item sets
 //Stpe 4–7: Mine k outliers data with minimum MFPOF value based on the obtained MFPs
 Step 4: foreach t in D
 According to formula (3), calculates the maximum frequent patterns factor of each record t: MFPOF(t);
 end foreach//Calculating maximum frequent factor of each transaction t
 Step 5: Obtains a MFPOF value of each network behaviour records t;
 Step 6: For all t, they are sorted in ascending order according to MFPOF (t);
 Step 7: Return the first k network behaviour record with the minimum MFPOF value, and they are k outlier data in the network behaviour data.
3.4. Automatically Constructing Intrusion Detection Patterns Based on Association

Association analysis can automatically discover the data characteristics of network behaviour. The maximum frequent patterns generated by association analysis can reflect the maximum common characteristics of network behaviour data, which are expressed by the attribute values of network behaviour data. So, these attribute values can be used to build intrusion detection patterns with strong classification ability [22].

Taking the outlier dataset obtained by MFPOF-OM algorithm as input and setting a minimum support threshold, the maximum frequent patterns of the outlier dataset can be obtained referring to Step 1–3 of Algorithm 1, which are the intrusion detection patterns of network attack.

3.5. System Architecture

According to the abovementioned analysis, the architecture of the system proposed in this work consists of six major modules such as data preprocessing, an outlier mining module, constructing intrusion detection patterns, attack patterns base, pattern match, and an alarm system, as shown in Figure 1.

The data preprocessing module is for performing preprocessing activities, but its main function is to discretize the data and make it suitable for the proposed algorithm. The outlier mining module is used to obtain the outlier data by the proposed algorithm. On the basis of acquiring outlier data, an intrusion detection pattern module is used to obtain intrusion detection patterns, so as to construct the attack pattern library module. The pattern match module is used to match the testing data with the attack rule base. If the match is successful, it indicates that there is an intrusion attack and transfers to the alarm module to trigger the alarm.

4. Results and Discussion

4.1. Dataset and Experimental Environment

The specifications of the hosts adopted in the experiments are Core Intel Core i5-6300HQ, 2.3 GHz CPU, 16 GB RAM, and Windows 7. The proposed method is verified in MATLAB 2012. The NSL-KDD dataset [23] and UNSW-NB 15 dataset [24] are used as the experimental datasets to verify the proposed method in this paper.

First, the experimental results of the proposed algorithm are analysed in the NSL-KDD dataset, and then, the proposed algorithm is compared with other researchers’ algorithms to verify the effectiveness it; lastly, the experimental results in the NSL-KDD dataset and UNSW-NB 15 dataset are compared to verify the applicability of the proposed algorithm.

The NSL-KDD dataset is an effective benchmark dataset to help researchers compare different intrusion detection methods. There are 125,973 connection records in the NSL-KDD dataset. Each connection record is described by 41 attributes about the network packet, network traffic, host traffic, and content information. The 22 categories of attacks are from the following four classes: DoS, R2L, U2R, and Probing. Also, the 20th attribute (num_outbound_files) can be deleted because its attribute value is all 0, so its information entropy is 0 according to information theory.

The raw network packets of the UNSW-NB15 dataset are created for generating a hybrid of real modern normal activities and synthetic contemporary attack behaviours. It is suitable for researchers to study the intrusion detection system. There are 175,341 records in the training set and 82,332 records in the testing set. This dataset has totally 49 features with the class label and 9 families of attacks, namely, Fuzzers, Analysis, Backdoors, DoS, Exploits, Generic, Reconnaissance, Shellcode, and Worms.

The NSL-KDD dataset is a factual benchmark in the field of network intrusion detection, which lays a foundation for the research of network intrusion detection based on computational intelligence. First, the NSL-KDD dataset eliminates duplicate records and classifiers that prefer more duplicate records. Second, it eliminates the imbalance between the number of records and reduces the false positive rate. Therefore, although the NSL-KDD dataset is older, it is widely used to evaluate the performance of the IDS. The UNSW_NB15 dataset is a comprehensive network attack traffic dataset, which combines the real normal network traffic attack activities and modern network traffic comprehensive attack activities and can better reflect the real environment of the network, so it is widely used in abnormal intrusion detection [25, 26].

The proposed algorithm needs to mine the maximum frequent pattern, which requires that the data type must be discrete. Taking the NSL-KDD dataset as an example, the dataset values’ processing is introduced, which is suitable for the proposed algorithm. According to the analysis of the NSL-KDD dataset, the attribute data type of the dataset can be divided into the text type and numerical type, and the numerical type can be divided into the discrete type and continuous type. The types of data are shown in Table 1 for the text-type and numerical discrete-type data which have met the data requirements. However, the continuous numerical data represented by columns 1, 5, and 6 are discretized using the discretization algorithm given in Section 3.2 and transformed into reliable and accurate data suitable for data mining.

4.2. Experiments in the NSL-KDD Dataset

Experiment A: the experimental results of the proposed algorithm in the NSL-KDD dataset are analysed in the experiment. The accuracy, false positive rate, and complexity analysis are used as the performance evaluation criteria to determine the results. Four groups of sample data were extracted from the dataset: Normal + DoS, Normal + Probing, Normal + R2L, and Normal + U2R.

4.2.1. Experiment Results of Four Network Attack Patterns

By comparing the detection rate and false positive rate under different MinSP thresholds of four groups of sample data, Normal + DoS, Normal + Probing, Normal + R2L, and Normal + U2R, the detection effect of the proposed algorithm is illustrated, and then, the feasibility of the proposed algorithm is verified. The experimental results of DoS, Probing, R2L, and U2R intrusion detection patterns obtained from the analysis of four groups of sample data are shown in Figure 2.

Probing attack detection patterns are taken as an example for data analysis. The Normal + Probing sample set contains 62000 pieces of data, the threshold value of MinSP is different, and the detection patterns are also different in the experiment. The experimental results are shown in Figure 2(b), which shows the detection patterns acquired under the MinSP thresholds of 58500, 59000, and 60000 and uses the acquired Probing detection patterns to detect five data types (DoS, Probing, R2L, U2R attack data, and Normal data), respectively. It is found that when the threshold value is 59000, the accuracy of Probing detection patterns to Probing data is 88%, and the false alarm rate is 2% to Normal data, 4% to DoS, 1% to R2L, and 10% to U2R data. When the threshold values are 58000 and 60000, the results are as shown in Figure 2(b) and will not be described one by one.

By comparing the four intrusion detection attack modes in Figure 2, it is found that the accuracy will be better when the minimum support threshold is larger, and the detection error of other data is basically the same, although the size varies. It is determined by the characteristics of outlier mining. The larger the threshold is, the fewer the number of outliers is, which can better reflect the characteristics of attack-type data. Of course, the threshold should not be too large, and the accuracy will be reduced if the threshold is too large. Through the comprehensive analysis of detection rate and false detection rate under multiple thresholds, the intrusion detection mode with the best comprehensive detection result is selected as the acquired intrusion detection mode, and the threshold value at this time is taken as the acquired intrusion detection pattern threshold: the threshold of DoS attack is 59100, the threshold of Probing attack is 59000, the threshold of R2L attack is 59600, and the threshold of U2R attack is 59500. The evaluation parameters are shown in Table 2.

Comparing the four subgraphs in Figure 2, it is found that U2R-type data have the highest detection errors in DoS, Probing, and R2L attack intrusion detection patterns, which are 4%, 10%, and 33%, respectively, and compared with the other three attack intrusion detection patterns, the accuracy of U2R attack intrusion detection mode is relatively low, only 87%, which is determined by the number of U2R, only 52 pieces of U2R data in the NSL-KDD dataset, so data mining cannot fully discover its data characteristics, resulting in incomplete detection performance.

Comparing Figure 2(c) with Figure 2(d), it is found that there are higher errors in the detection of U2R data by using R2L attack intrusion detection patterns and R2L data by using U2R attack intrusion detection patterns, which shows that R2L-type data and U2R-type data have higher data similarity compared with other three types of data, which is consistent with the characteristics of two kinds of network attacks in reality.

4.2.2. Complexity Analysis

In this section, the complexity of 4 groups of sample data, Normal + DoS, Normal + Probing, Normal + R2L, and Normal + U2R, will be analysed. The FindFPOF algorithm based on frequent patterns and other outlier mining algorithms based on weighted frequent patterns need to mine frequent patterns first, and the time complexity is similar. Here, FindFPOF algorithm is taken as an example to illustrate.

The total time complexity of FindFPOF algorithm is O (m2 + mn + mlogm), where m is the amount of data and n is the amount of frequent patterns.

The MFPF-OM algorithm has three steps: (1) mining maximum frequent patterns from the dataset, the time complexity is O (m2); (2) calculating the MFPOF(t) of each network behaviour record, the time complexity is O (ml); and (3) discovering K network behaviour outliers, the time complexity is O (mlogm). Therefore, the time complexity from the abovementioned three steps is proved as follows: T(MFPOF-OM) = O (m2 + ml + mlogm), where m is the number of data and l is the number of maximum frequent patterns.

The number of frequent patterns (n) in FindFPOF algorithm and the number of maximum frequent patterns (l) in MFPF-OM algorithm for 4 groups of sample are shown in Table 3.

For massive data, the value of m is large enough, and in theory, the time complexity of the two algorithms can be simplified to O (m2). But in practice, when the value of m is not large enough, the proposed algorithm only needs to mine the maximum frequent patterns in Step 3, and l ≪ n, as shown in Table 3, so MFPOF-OM algorithm has a better time complexity than the FindFPOF algorithm when calculating MFPOF (t) in Step 4 of the algorithm.

4.3. Comparative Experiments between the Proposed Algorithm and Other Algorithms

Experiment B: in order to verify the accuracy of the proposed method, it is compared with the SVM method, Intelligent DT method [6], LSSVM + FRFSA method [5], and Outlier Detection + EMSVW method [20]. The accuracy is used as the performance evaluation criteria to determine the results. The evaluation parameters are shown in Table 4.

The results are shown in Figure 3, in which M1 represents the SVM method, M2 represents the Intelligent DT method, m3 represents the LSSVM + FRFSA method, M4 represents the Outlier Detection + EMSVW method, and M5 represents the proposed method in this paper. The results show that the MFPOF-OM method is very close to the other methods in accuracy of Probing and DoS, but slightly inferior. However, it has a great advantage in the accuracy of R2L and U2R, which shows that the improved dimensional outlier mining method has good characteristics in dealing with outlier data because of the small amount of R2L and U2R attack data in the NSL-KDD dataset. The accuracy data of R2L and U2R are empty in Figure 3 because there are no relevant data in [20]. The overall performance analysis shows that the performance of the proposed method is reliable, can effectively detect the intrusion behaviour in network data, and can meet the actual operation requirements.

4.4. Comparative Experiments between the NSL-KDD Dataset and UNSW-NB15 Dataset

Experiment C: in this experiment, the proposed method is tested and compared in the NSL-KDD dataset and UNSW-NB15 dataset, and the performance of the proposed algorithm is estimated by using the performance metrics, namely, precision, recall, and F1-measure and ROC. The two datasets have different attack patterns and data characteristics, so it is impossible to compare each pattern separately, and only the overall performance index is analysed in two datasets in this paper. The overall performances of precision, recall, and F1-measure in the two databases are shown in Table 5. Figure 4 shows the comparison results of precision, recall, and F1-measure in two different databases.

Figure 5 shows the ROC curves in two different databases. It is found that although the detection results of the UNSW-NB15 dataset are better than those of the NSL-KDD dataset in some values, the detection results of the NSL-KDD dataset are generally better than those of the UNSW-NB15 dataset from the whole ROC curve.

By comprehensively comparing the performance indexes in Figures 4 and 5, it is found that the proposed method’s technique achieves better performances for the NSL-KDD dataset. The reason is that some malicious records in the UNSW-NB15 one are not high because of the lower variances between them and normal records, and the data are optimized in the NSL-KDD database, which is more suitable for the detection of malicious records. But on the whole, it shows very good performance in the NSL-KDD dataset and UNSW-NB15 dataset, which proves the effectiveness of the proposed method in high-dimensional anomaly detection.

5. Conclusions

In this paper, a high-dimensional outlier mining algorithm based on the maximum frequent pattern factor (MFPOF-OM) has been proposed by using the related technology of high-dimensional outlier mining based on frequent patterns. This work has two advantages: first, the MFPOF-OM algorithm only needs to mine the maximum frequent pattern set, which solves the problem of mining completely frequent patterns in frequent pattern outlier algorithm; second, it can greatly reduce the number of maximum frequent patterns, thus reducing the time complexity of the algorithm. Experimental results show that the proposed method is feasible, which can further reduce the time complexity while ensuring the excellent detection performance compared with the contrast algorithms.

Data Availability

The data used to support the findings of this study are included within the article.

Conflicts of Interest

The authors declare that there are no conflicts of interest regarding the publication of this paper.


This work was supported by the National Natural Science Foundation of China (No. 61772450) and Hebei Province Natural Science Foundation of China (F2017203307).