AIRCC PUBLISHING CORPORATION
Multi-Layer Classifier For Minimizing False Intrusion
Shaker El-Sappagh, Ahmed saad Mohammed, Tarek Ahmed AlSheshtawy
Faculty of Computers & Informatics, Benha University, Egypt.
Intrusion detection is one of the standard stages to protect computers in network security framework from several attacks. False alarms problem is critical in intrusion detection, which motivates many researchers to discover methods to minify false alarms. This paper proposes a procedure for classifying the type of intrusion according to multi-operations and multi-layer classifier for handling false alarms in intrusion detection. The proposed system is tested using on KDDcup99 benchmark. The performance showed that results obtained from three consequent classifiers are better than a single classifier. The accuracy reached 98% based on 25 features instead of using all features of KDDCup99 dataset.
Intrusion detection, multi-layer classifier, KDD CUP 99, False Alarms
Nowadays, communication technology is widely used for communication and transmission purposes in many applications working on different platform. This captures the attention towards network security. In this technology, the security becomes a challenging problem and vulnerable for intrusions. Subsequently, it is necessary to utilize a system for network security such as the intrusion detection systems (IDSs) [1, 2]. Network Intrusion Detection systems had become the most important components of recent network infrastructure due to the effects of fast security threats in today’s computer network. Intrusion detection system is generating a good number of alarms; however, it deployed algorithmic procedures to reduce false positives . The great number of false positive alarms is made it difficult for security analyst to recognize successful attacks and to take reconditioned actions for such threat. Alarms contain a high rate of unsuccessful alarms recognized as false alarms which demand estimation to recognize and reduce the unsuccessful ones. These are raised once an intrusive incident detected and presenting the security analyst the chance to react against any such threat . Data Mining (DM) and machine leaning are commonly utilized within the scope of intrusion detection to find the hidden patterns of intrusions and their connection surrounded by each other . DM can be applied to understand from traffic data utilizing the target learning procedures to discover intrusion models of evaluation learning procedures to recognize dubious actions .
Classification is one of DM categorization that assigns topics to one of several classes .The huge issue in IDSs that the presence of huge number of false alarms; this issue motivate several experts to discover the solution for minifying false alarms.This research will be suggesting multi-classifier for handling this issue.
In this research, a new multi-classifier is suggested for improving the minifying false alarms in intrusion detection system. The basic idea is to use multi classifiers instead of using one classifier. The remaining of this paper is organized as follows: In section 2, we present previous works of applying classification techniques for intrusion detection. Section 3, we present the description of KDDCUP99 dataset. Section 4, we discuss the attacks type of KDDCUP99. Section 5, we display the performance measure of the intrusion detection system. Section 6, we explain all steps of the proposed system, Section 7, we display the results that are obtained from the proposed system and finally in Section 8, we provide the important conclusions of the proposed system
2. Previous Work
Many classification techniques are used for classifying intrusion detection datasets. Classifiers have been suggested and developed to reduce false alarm of intrusion detection in the scope of network security based on different ideas. Recently papers in this scope can be summarized as follows: Aggarwal and Sharma , selected a set of classification algorithms on the basis of their effectiveness according to speed and, the ability of handling large data-set and after that, for KDD’99 data-set simulated 10 selected existing classifiers according to Weka tool. Manju , analyzed the performance of the Intrusion Detection System using various classification approaches. The objective of this paper was to analyze and predict the network attacks by classifying them as normal and abnormal, and for implementing and measuring the efficiency of this system, the standard KDD99 benchmark data-set has been chosen. Gupta. et al., implemented a variety of data mining procedures which consisted of linear regression and K-means for automatic generation of rules for classifying network activities. A comparative study of those algorithms for detecting intrusions has been made based on the KDD99 data-set as well. Liyu Duan and Youan Xiao  large volume of the data and unbalanced data ,intrusion data were inevitable obstacles. So, to solve those issues utilizing fuzzy c-means procedure to reconstruct feature vectors according to central points. This paper shows deep learning procedure for intrusion detection, which implemented in GPU-enabled TensorFlow and evaluated utilizing the KDD 99 dataset. Osamah et al  in his paper, introduced learning procedure for intrusion detection according to tree calculation on the KDD-99.
3. KDDCup99 dataset
In order to apply the proposed mechanism, the KDDCup99 dataset will be used as the standard dataset. KDDcup99 data-set has been considered the point attraction for numerous researchers in the domain of intrusion detection system. It has been most widely used for evaluating IDS. The 10% of KDDCup99 dataset is the original data-set that includes 494,020 records as showed in Figure 1 .Every record includes forty-one features with either normal or abnormal class with one fixed attack such as Dos, Probe, U2r, R2l as shown in Table 1 ,.
Figure 1. Sample of 10% Kddcup99 in Text Format
Table 1: Num of Records and Attack category of KDD’99 Dataset
Features of KDD’99 Dataset can be labeled into following:
Table 2: The No. Attributes of KDDCUP99
4. Attacks Type Of Kddcup99
Attacks occur into one of the following types :
The performance of classifier can be estimated by using various procedures according to the following criteria :
Table3: Confusion Matrix
FAR= FP/ (TN+FP) (1)
Specificity = TN/ (FP+TN) (3)
6. The Proposed System
Multi-layer classifier can be comprised into three layers. The first layer will detect the abnormal from normal traffic using naïve Bayes, the second layer will classify the abnormal activity into 4 classes of attack and normal activity using neural network (backpropgaion), while the last one is to classify 4 classes of attack into 23 subclass and normal activity using Decision Tree. The outcomes of the Multi-layer classifier are evaluated in testing stages. The general structure of the Multi-layer classifier of classifying intrusion detection is demonstrated in details as shown in figure 2.
The proposed of multi-layer classifier is training and testing based on KDDcup99 dataset, this work have been used the whole dataset about 494,020 records which include normal behavior samples and attack types. Selected training data is about 329,510 records while the testing data is 164,510. To evaluate the proposed system cross validation will be used by splitting the training and testing data in k times. Table 4 displays the type of classes, which are used in dataset.
Fig 2: General Structure of the Multi-layer classifier
Table 4: Description of Category Traffic Dataset
6.1 The Proposed Procedures For Classifying Intrusion
The proposed procedure for classifying intrusion on KDDcup99 dataset is shown in algorithm 1 according to multi-operations and multi-layer classifier. Multi-operations can be comprised into the following steps:
The Dataset includes 41 features with a label to determine the type of each record whether normal or type of attack traffic. In addition to the main class attribute of 23 subclass types in dataset, two features of class categories have been added to the data and reach up to 44 attributes according to the 23 subclass to their main class type of normal and abnormal traffic as shown in Table 2. These features are treated as class category which are used in the experiments of the system. Kddcup99 dataset will be split randomly into two non-overlapped parts 329,510 records of training data and 164,510 records of testing data. According to the cross validation procedure, the data is splitted into 3 equal bins of folds approximately. Training and testing data are performed in k times. In first layer, fold 1 is reserved for the test data and the remaining folds (fold2, fold3) are used to train using Naïve bays classifier. The second layer, fold 2 is reserved for the test data and the remaining folds (fold1, fold3) are used to train using backpropgaion classifier , the third layer, fold 3 is reserved for the test data and the remaining folds (fold1, fold2) are used to train using ID3. Evaluation metric in terms of accuracy are used for evaluating the efficiency of the suggested system.
The experiments showed that after examining the results of the data, it was found that the accuracy of a multi-layer classifier with three classifiers as shown in Table 5, Table 6 and Table 7 was better than the accuracy by using one classifier. If one classifier as ID3 was adopted as a single part, it does not give a precision index and the second one of Naïve Bayes classifier. Especially, in the case of equal values in the results or assigned number after the interval, therefore it will use more than one level to give a better decision from a single classifier where it increases accuracy, especially applications that rely on accuracy such as intrusions.
Table 5: Total Accuracy for First Layer
Table 6: Total Accuracy for Second Layer
The performance of proposed procedure showed that the total accuracy in the first layer selecting 30 features had best results, the total accuracy in second layer and third layer selecting 25 features had best results as described in Figures 3.
Fig 3: Average Accuracy of Three Layers
The result of the first layer, second layer and third layer were compared and evaluated with criteria FAR, specificity, sensitivity and time in minutes for 25 features as shown in Table 8. Table 9 displays the comparison of experimental result between the prior studies and proposed system.
Table 8: FAR, Specificity, Sensitivity and Time for Multi-classifier=
Table 9: Comparison of Related Work and Proposed System.
False alarms are a big problem in intrusion detection.This paper handled false alarms in intrusion detection by presenting a proposed multi-operations and multi-layer classifier. Depending on the results of the proposed approach to classify the type of intrusion and classifying them into their subclasses of intrusion. There are several conclusions from our study, The use of supervised machine learning classiﬁers such as Naive Bayes, ANN, and decision tree give the high efficiency and accuracy for the proposed approach. Using cross validation technique estimates and compares the performance of different algorithms and finds the best one from available data. Since it is very large dataset, we applied the cross validation technique to avoid falling into over fitting. The performance showed that results obtained from three consequent classifiers are better than single classifier.The accuracy reached 98% based on 25 features instead of using all features of KDDCup99 dataset.
 Vaidya h., Mirza SH., and Mail N.,”Intrusion System”, International Journal of advance research in engineering, science and technology, e-ISSN:2393-9877, p-ISSN:2394-2444,vol 3, Issue 3, Mar 2016.
 Nadiammai G.V and Hemalatha M.,”Effective Approach Toward Intrusion Detection System Using Data Mining Techniques”, Elsevier B.V. Egyption Informatics Journal, 2014.
 Islam A. and Islam M., “A Novel Signature_Based Traffic Classification Engine Reduce False Alarms in Intrusion Detection systems”, International Journal of Computer Networks and Communications (IJCNC) vol 7, No.1, Jan 2015.
 Al-Saedi K. and Manickam S., “research proposal: An Intrusion Detection System Alert Reduction and Assessment Framework Based on Data Mining”, Journal of Computer Science, 9(4):421-426,2013.
 Dult, I. and Dr.Borah S. “Some Studies in The Intrusion Detection Using Data Mining Techniques”. International Journal of Innovative Research in Science, Engineering and Technology, 4(7), 2015.
 Novakovic, J., Strbac P. and Bulatovic, “Toward Optimal Feature Selection Using Ranking Methods and Classifications Algorithms”. Yugoslav Journal of operations research, 2011.
Goeschel K., “Reducing False positives in Intrusion Detection systems using Data-Mining Techniques utilizing Support Vector Machines, Decision Trees and Naïve Bayes for off-line analysis”, IEEE, International Conference on 30 March-3 April 2016, USA, 7506774, July 2016.
 Mahmood D.Y., “Classification Trees with Logistic Regression Functions for Network Based Intrusion Detection System”, (IOSR-JCE) Journal of computer Engineering, e-ISSN: 2278-0661, p-ISSN: 2278-8727, vol 19, Issue 3, pp 48-52, June 2017.
Belavagi M.C. and Muniyal B., “Performance Evaluation of Supervised Machine Learning Algorithms for Intrusion Detection”, (IMCIP) International Multi-conference on Information Processing-2016, Elsevier, vol 89, pages 117-123, 2016.
 L. Duan and Y. Xiao, “An Intrusion Detection Model Based on Fuzzy C-means Algorithm,”, 8th International Conference on Electronics Information and Emergency Communication (ICEIEC), Beijing, pp. 120-123. (2018)
 Wang, Zheng. “Deep learning-based intrusion detection with adversaries.” IEEE Access 6 , 38367-38384.(2018):
Raheem Esraa and Saleh Alomari ,”An Adaptive Intrusion Detection System by using Decision Tree Osamah Adil”, Journal of AL-Qadisiyah for computer science and mathematics Vol.10 No.2,(2018).
Siddiqui M.K. and Naahid Sh., “Analysis of Kdd cup 99 Dataset Using Clustering Based Data Mining”, International Journal of database theory and application, pp.23-34, vol.6, No.5(2013).
Tavallaee M., Bagheri E., Lu W. and Ghorbani A.A, “A Detailed Analysis of the Kdd Cup 99 Data set”, proceedings of the IEE symposium on computational intelligence in security and defense applications ,2009.
Nelcileno A., R. Oliveira, A. Akira Shinoda, B. Bhargava, “Identifying Important Characteristics in the KDD99 Intrusion Detection Dataset by Feature Selection using a Hybrid Approach”, pp:2, 2010.
 Mahbod T., E. Bagheri, Wei Lu, and A. A. Ghorbani ,” A Detailed Analysis of the KDD CUP 99 Data Set”, p.2, 2009.
 Brifcani A.M.A. and Issa A.S., “Intrusion Detection and Attack Classifier based on Three Techniques: A Comparative Study”, Journal of engineering and Technology, Vol.29, No.2, 2011.