Vous êtes sur la page 1sur 7

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING 1

A Novel Intrusion Detection System for


RPL-Based Cyber–Physical Systems
Un nouveau système de détection d’intrusion pour
les systèmes cyber-physiques basés sur RPL
Mridula Sharma , Haytham Elmiligi, and Fayez Gebali , Life Senior Member, IEEE

Abstract— The physical layer of cyber–physical systems (CPSs) is composed of resource-constrained devices
connected in a wireless sensor network (WSN). Although this layer is easy to deploy, in most cases, it has
many security issues. Several intrusion detection systems (IDSs) have been proposed and tested as effective and
efficient solutions to detect only a few known attacks. In this article, we propose a novel, Supervised machine
learning-based IDS that is capable of detecting several attacks. This article discusses all IDS design steps, starting
from data collection to the feature engineering analysis and building the trained models. Experimental results
show that the proposed IDS can detect four different types of attacks that were seen by the machine learning
models during the training phase. The IDS can also detect the existence of several other attacks that are not seen
by the model and classify them as unknown attack types. The proposed model achieves 99.97% classification
accuracy when detecting known attacks and 85% classification accuracy when detecting a new attack type.
Résumé— La couche physique des systèmes cyber-physiques (CPS) est composée de dispositifs à ressources
limitées connectés dans un réseau de capteurs sans fil (WSN). Bien que cette couche soit facile à déployer,
dans la plupart des cas, elle présente de nombreux problèmes de sécurité. Plusieurs systèmes de détection
d’intrusion (IDS) ont été proposés et testés comme solutions efficaces et efficientes pour détecter seulement
quelques attaques connues. Dans cet article, nous proposons un nouveau système de détection d’intrusion
basé sur l’apprentissage automatique supervisé, capable de détecter plusieurs attaques. Cet article aborde
toutes les étapes de la conception de l’IDS, depuis la collecte des données jusqu’à l’analyse de l’ingénierie des
caractéristiques et la construction des modèles formés. Les résultats expérimentaux montrent que l’IDS proposé
peut détecter quatre types d’attaques différentes qui ont été vues par les modèles d’apprentissage automatique
pendant la phase de formation. L’IDS peut également détecter l’existence de plusieurs autres attaques qui ne
sont pas vues par le modèle et les classer comme des types d’attaques inconnus. Le modèle proposé atteint une
précision de classification de 99,97 % lors de la détection d’attaques connues et de 85 % lors de la détection
d’un nouveau type d’attaque.
Index Terms— Correlation-based feature selection, cyber–physical system (CPS), decision tree classifier (DTC),
embedded method, feature engineering, feature reduction, filter method, information gain, logistic regression
classifier (LRC), naive Bayes classifier, random forest classifier (RFC), routing protocol for low power and lossy
networks (RPLs), state vector machine (SVM), wireless sensor network (WSN) security, WSN attacks.

I. I NTRODUCTION traffic for further processing to the cyber layer [2]–[4].

T HE physical layer of cyber–physical systems (CPSs)


is composed of resource-constrained devices connected
in a wireless sensor network (WSN). routing Protocol for
Multihop routing methodology for transmitting data from
sensor nodes to border routers makes the network very much
prone to routing attacks, such as hello flood, decreased rank,
low-power and lossy networks (RPLs) is one of the most increased version, selective forwarding, and DDoS attacks [2],
important protocols used at this layer for both static networks [5]. These attacks happen at the physical layer of the CPSs
(such as BAN and smart grid) and dynamic networks (such as the physical layer runs on the RPL protocol.
as vehicular Area Network) [1]. It provides a mechanism for Several intrusion detection systems (IDSs) have been pro-
point to multipoint traffic (P2MP) and multipoint-to-point posed in the literature for the RPL network [6]–[8]. Packet
traffic (MP2P). The traffic flows from 6LoWPAN nodes (i.e., compression headers have been used to analyze the network
from the end or router nodes) to the sink node or border traffic in order to design an IDS, known as CHA-IDS [8].
router, which is the central control point, which transmits Other examples of IDS solutions for RPL include SVELTE and
Pongle’s that are useful for detecting sinkhole and wormhole
Manuscript received October 13, 2020; accepted January 11, 2021.
(Corresponding author: Mridula Sharma.) attacks, respectively [6], [9]. CHA-IDS detects only three
The authors are with the Department of Electrical and Computer Engi- attacks, namely, hello flood, sinkhole, and wormhole, whereas
neering, University of Victoria, Victoria, BC V8P 5C2, Canada (e-mail: SVELTE and Pongle are useful for detecting sinkhole and
naina@uvic.ca; haytham@ieee.org; fayez@uvic.ca).
Associate Editor managing this article’s review: Aiman Erbad. wormhole attacks only. Although these IDS solutions are all
Digital Object Identifier 10.1109/ICJECE.2021.3053231 efficient in detecting these specific attacks, they are not capable
2694-1783 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: East Carolina University. Downloaded on June 15,2021 at 16:42:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING

of detecting any new attacks. After reviewing several articles TABLE I


and studies in the literature, a few findings are as follows. S UMMARY OF A LGORITHMS U SED AT T HREE L AYERS
1) Almost all of these IDS are only effective for specific OF THE P REDICTIVE M ODEL
attack types and cannot detect multiple or combinations
of attacks. They are also unable to detect a brand new
attack.
2) Almost in all the works, IDS has only undergone n-fold
cross-validation testing on the data set. None of the
models are tested using new data that are not seen by
the model in the training phase.
3) KDD and many other openly available data sets are com- the normal network traffic and instances when the network
monly used in IDS research. The applicability becomes is under attack. Data are collected from Cooja simulator log
questionable as these data sets are not essentially for files and are converted into a form that is easier to interpret by
WSN scripts that train the model. Following that, feature engineering
These outcomes provide us with sufficient motivation to do is done, which includes both feature generation and feature
our research to develop an IDS that can detect known attacks reduction to get an optimal feature subset.
and new attacks as well. 1) Feature Generation: Using the pcap capture of each run,
Our proposed IDS addresses this problem and provides a UDP data are filtered from the rest of the data in two separate
mechanism to detect both known attacks found in the training XML files using command line parser tShark, as shown in
and several unseen/unknown attacks. Our model is further Algorithm 1. Following that, each XML file is parsed using
extended to detect any suspicious activity in the network that a custom script to generate selected data in the.csv file.
does not conform to the normal traffic flow. This process is explained in Algorithm 2 for UDP data and
This article has the following main contributions. Algorithm 3 for all data. Finally, all generated.csv files are
1) Proposed a complete packet analysis model of RPL processed and analyzed to generate a final list of features,
protocol. The analysis identifies the features that can be as shown in Table II.
used to identify various traffic patterns under different A test environment is created in a Cooja simulator to
circumstances in CPSs. generate the data used to build the training set. Once each
2) Built a new data set based on 300 different simulations packet of one sample run is parsed, a summary of the complete
of RPL networks under different normal and attack simulation is generated as one row of data in the final data set.
scenarios. Based on the scenario in the simulation, a class label is
3) Constructed a new IDS framework for RPL networks. assigned to every row, which represents the type of network
4) Proposed a novel predictive model for intrusion detec- traffic of that specific run. If the network has no attack, then
tion for four known attacks using machine learning the class label is set to 0 indicating no attack. The other four
analysis. An extension of the predictive model is added attack scenarios that are simulated are given labels as 1–4,
so that it can detect a new unseen attack or a combina- indicating Hello Flood attack, DIS attack, Increased version
tion of several attacks. attack, and Decreased rank attack, respectively. As a result of
The rest of this article is organized as follows. Section II this step, a data set comprising 58 features (except class label)
explains the IDS framework that is used to build the machine is created. Table II lists the features and their descriptions. The
learning model. Section III discusses the experimental results, generated data set represents the traffic patterns of four attack
using both n-fold cross validation and validation using unseen vectors in addition to the normal network traffic.
data. Finally, Section IV concludes this article and proposes To make an efficient and successful IDS, dimensionality
new directions for future work. reduction is required to reduce the overall time and processing
II. P ROPOSED A RCHITECTURE complexity of the implementation. Moreover, all 58 features
The proposed IDS is a complete framework, as shown are not equally important and do not contribute the same
in Fig. 1. weight to the classification accuracy of the IDS system.
There are three layers of the framework. Therefore, feature reduction plays a vital role in improving
1) Layer 1: Data set building. the overall performance of the system.
2) Layer 2 Model building & evaluation. 2) Feature Selection/Reduction: The generated data set
3) Layer 3: Predictor. includes a list of 58 features, as shown in Table II. These
features are saved in comma-separated value (.csv) format file.
As shown in the figure, each layer has different roles, and
However, some of this information is not needed during the
for each role, one or more algorithms are combined together.
training phase as it creates pattern overlaps between classes
Table I lists the tasks done at each layer of the model.
and may lead to incorrect classification in some cases. In addi-
In the following, the three layers of the framework are
tion, having all 58 features will require several processing
explained in detail.
cycles on the target sensor motes.
A. Layer I: Data Sensor Feature reduction is required to select only the most signif-
A supervised machine learning model is trained using a icant features that can be used to distinguish different attack
training data set. The training data set contains instances of patterns. This step also helps the machine learning algorithm to

Authorized licensed use limited to: East Carolina University. Downloaded on June 15,2021 at 16:42:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SHARMA et al.: NOVEL IDS FOR RPL-BASED CPSs 3

Fig. 1. IDS Framework.

TABLE II Five different models are trained using the training set, and
D ATA S ET W ITH THE C OLLECTED F EATURES their accuracy scores are collected. After feature reduction,
the accuracy scores of each classifier need to be compared
to see the overall impact of the feature reduction method.
We have used two different methods for feature reduction so
that we may compare the results to make an optimized model.
The two methods are as follows.
Calculate the Importance of Each Feature Using RFC
(Embedded Method): Embedded methods use a subset of
features and train a model using them. Based on the results
from the model, features are added or removed. We used
random forests classifier to train and validate. Random forests
are very commonly used for feature reduction, as it uses
tree-based strategies that naturally rank the features. Nodes
with the greatest importance happen to be placed at the start
of the trees, while the less important features are at the end
of the trees. Thus, the pruning of the trees below a particular
value can generate a subset of the most important features.
The accuracy score using features identified with different
thresholds is tested. Starting with the threshold > 0.020,
which selected 16 features, the next threshold value of
>0.030 yielded 11 features, a value of >0.40 selected eight
get trained faster and reduces the complexity of the generated
features, and a value of >0.50 selected only four features.
model.
Initially, when the features have been selected, it is a compu-
Five different machine learning classifiers are used for
tationally expensive method, but, later, it helps in saving the
model building. These classifiers are random forest classifier
prediction time considerably, as can be seen in Section III. The
(RFC), state vector machine (SVM), decision tree classifier
final feature set of eight features is defined with the features
(DTC), naive Bayes classifier (GNB), and logistic regression
having importance >0.04, as explained in Algorithm 5.
classifier (LRC).
Correlation-Based Feature Selection (Filter Method): This
These classifiers are chosen for two reasons.
method selects the features that are highly correlated with the
1) They have been proven effective for similar classification class, but uncorrelated with each other [11]. Here, the model
problems [10]. accuracies are again tested with the reduced features. The
2) These classifiers are not very computationally intensive, accuracy score uses features that are identified with dif-
i.e., need low resources to process. ferent correlations. Starting with the correlation of >0.25,

Authorized licensed use limited to: East Carolina University. Downloaded on June 15,2021 at 16:42:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING

which selected 12 features, the next correlation value of 1) a normal network with no attacks;
>0.30 selected ten features, and then, >0.35 selected only 2) a network with one of the four known attacks;
five features. The improvements both in accuracy and time 3) a network with a new attack or a combination of several
took for training and prediction are discussed in Section III. attacks.
The final feature set of ten features is used where features This predictive model is capable of identifying any of the
have a correlation > 0.30, as explained in Algorithm 6. The abovementioned scenarios, using the process defined in the
five models were retrained using the new feature set, and it following.
showed improvements both in accuracy and time taken for 1) Normal Traffic: If three or more classifiers identify
training and prediction, as explained in Section III. class 0 for this testing set, the network undoubtedly does
not have an intrusion and the system continues working as
B. Layer 2: Model Building & Evaluation before.
Using the optimal feature set, the five models are trained. 2) Identifying a Known Attack: If three or more classifiers
To improve the success rate of the model, three steps were identify classes 1–4, a known attack seems to occur in the
taken. network. An alarm is raised to the administrator of the network
1) A set of five classifiers is used to make predictions on showing a possibility of a known intrusion.
the attack scenario. 3) Identifying a New Attack: This scenario is where predic-
2) Feature reduction is done using two methods, i.e., the tion from different classifiers does not match. Since different
filter method and the embedded method, and selecting classifiers identify the intrusion differently, it indicates that an
the optimal feature set accordingly. attack is found in the network, which is not seen by the models
3) Making iterative predictions to make it more accurate if in the past in the training set. To make the prediction more
no clear results are predicted in the first attempt. accurate, the predictor performs 20 iterations to look for the
model results. If after 20 iterations, still no model, that is able
The two models built using a filter and embedded methods
to give a clear prediction, leads to the agreement of the new
are evaluated in two steps.
attack.
Step I (Model Evaluation Using n-Fold Cross Validation):
The n-fold cross validation is the common method of the
model evaluation. It works by splitting the data set into n-parts D. Predictor Deployment
(e.g., n = 5 or n = 10). Each split of the data is called a fold. The IDS model built is deployed on the border router.
The algorithm is trained on n − 1 folds with one held back A new.pcap file is collected from the network every 15 min
and is tested on the held back fold. This is repeated so that on the router. The features are extracted from the newly
each fold of the data set is given a chance to be the held back generated.pcap file using Algorithms 1–5 to generate the latest
test set. The classification accuracy, TP, TN, FP, FN, precision, testing set. This testing set is passed through the model for
and recall scores of all the selected algorithms are calculated. analysis. As the result of the analysis, if the predictions
The objective is to reduce the false alarm rate and increase the from different classifiers are the same, the result is clear.
detection rate. The test size of 0.2 indicates 20–80 test split There is always a possibility that different classifiers may
for the model evaluation. Algorithm 8 shows the procedure give different results. The final result declaration is through
used for n-fold cross validation. polling.
Step II (Model Evaluation Using New Data From the Similar Polling: The concept of selecting the majority decision
Network): For extended model evaluation, it is tested with is polling. If the majority is clear and most of the models
new data. The testing data are brand new data, which has declare 0 as result, there is no attack in the network. However,
never been seen by the model before. To do so, new.pcap if the polling gives the majority decision of 1–4, an intrusion
files are collected from the same network after running new is there and is declared based on the class. There may be an
simulations. The new.pcap file is used to generate an absolute instance when there is no clear majority decision. In that case,
new.csv file using algorithms (1 to 5). The new.csv data is 20 iterations are done for the prediction. If, in 20 iterations,
given to the predictive model for analysis. Since the predictive a single class is predicted for more than ten times (>50%
model has five submodels of five different classifiers, different times), the class value is used for the final declaration, but,
classifiers may give different predictions. For the purpose of if the majority is not very clear (i.e., is less than ten times),
making an accurate final prediction, polling is applied to the it is flagged as a new attack. This latest.csv is saved as an
results from five classifiers, and the majority class definition is archive file for future analysis. After the final prediction, if the
used as the final result. The polling results are also explained class predicted is 0, nothing is done, but, in all other cases,
in Section III. an alarm is raised to the administrator of the network about
the intrusion in the network.
C. Layer 3: Predictor
Now, our model is trained for the network with either no III. R ESULTS AND D ISCUSSION
attack or one of the four attacks, namely, hello flood, DIS The proposed IDS framework is divided into three layers.
attack, Increased version, or decreased rank. However, when Layer 1 includes feature engineering, and layer 2 executes
the model is deployed in the network, we may come across model building and evaluation algorithms. Layer 3 classifies
three different scenarios, which are as follows: the traffic and detects the attacks in the network.

Authorized licensed use limited to: East Carolina University. Downloaded on June 15,2021 at 16:42:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SHARMA et al.: NOVEL IDS FOR RPL-BASED CPSs 5

TABLE III TABLE IV


C HANGE IN THE C LASSIFICATION A CCURACY OF D IFFERENT TP S CORES OF X-VALIDATIONS U SING THE E MBEDDED M ETHOD
C LASSIFIERS W ITH C HANGING T HRESHOLD VALUES

TABLE V
TP S CORES OF X-VALIDATIONS U SING THE F ILTER M ETHOD
A. Layer 1: Feature Engineering
The.pcap file of each run is used to build a single instance
of the data set. Each instance has 58 features and one class
value. The data set has five different class values. For each
class, 60 instances are used to create the test data set. Two
different feature reduction methods are applied at layer 2.
1) Embedded Feature Reduction Method Using RFC: In
this method, we selected features with importance 1) n-Fold Cross Validation Results: For model testing,
> 0.05. Only eight features were found meeting the 80/20 cross validations are performed. The analysis for class 0,
threshold of 0.04. The experimental results show that, class 1, class 2, class 3, and class 4 is summarized in the
with only eight features, the performance of two out of following.
five classifiers improved, whereas the accuracy remained a) Results of the model built using embedded method:
constant for all other classifiers. The performance of the classifiers is evaluated against indi-
2) Correlation-Based Feature Reduction This method vidual scenarios.
reduced the features to 10. The experimental results In every case, TP and FP rates are evaluated. Table IV lists
show that the accuracy of one classifier improved, one the TP scores of models built using the embedded method.
dropped a little, and the other three remained almost b) Results of the model built using correlation (Filter)
constant. The final threshold chosen is 0.30. method: Correlation-based feature reduction removes features
1) Results of Embedded Method—Feature Reduction Using that have a low correlation with the class variable. The
Random Forest Classifier: In the embedded method, an RFC performance of the classifiers is evaluated against individual
is used to check the best couple of features with different scenarios.
thresholds and their impact on the prediction accuracy of the In every case, TP and FP rates are evaluated. Table V lists
model. We tested for the different values of threshold ranging the TP scores of the model built using the filter method.
from 0.02, 0.03, 0.04, and 0.05. The features were reduced to 2) Model Evaluation Results Using New Unseen Data:
16, 11, 8, and 4 based on these values. We finally selected Another evaluation method used is testing the model with the
the value of 0.04 that reduced the features to eight only (88% new data from the same network with the known class label.
feature reduction) as it improved the efficiency of SVM from To do so, new log data from the network are taken as a.pcap
63.33 to 83.33 and LRC from 97.77 to 98.88. GNB reduced file and are passed through Algorithms 1–5 to generate a new
from 80 to 75.55, but, for RFC and DTC, accuracies were data set in form of a.csv file. This file is passed to the model
constant, as listed in Table III. for prediction. All the five classifiers used for predictions may
2) Results of Filter Method—Correlation-Based Feature give the same or different predictions. The five results are
Reduction: Here, we tested the model accuracy with different used for polling, and the majority decision is taken as the
values of the correlation. The correlation values were set final result.
to greater than 0.25, 0.30, and 0.35 leading to the reduced 1) All Classifiers Give Same Result: The result is predicted.
features 12, 10, and 5. We chose the value of 0.30 as it reduced 2) Three or Four Classifiers Give Same Result: The major-
the features to 10 and improved the prediction accuracy of ity result is predicted again.
SVM from 63.33 to 77.77, keeping the rest almost the same. 3) Four Out of Five Classifiers Gives Different Results: In
This was a 77.5% feature reduction. this, the prediction analysis is repeated 20 times. To get
the final majority out of all runs, if the classification gets
B. Layer 2: Model Evaluation a rate of more than 50%, it is taken as the final result.
In our study, to make an accurately predictable IDS, we used For model testing with new data, ten simulations each were
five classifiers. After feature reduction using embedded meth- done for class 0, class 1, class 2, class 3, and class 4. The
ods and filter methods, two-way model evaluations are done, results found are summarized in the following.
which are n-fold cross validation and validation using new a) Results of the model built using embedded method:
data from a similar network. The results of the two evaluation Table VI lists the results of the model built using the embedded
methods are discussed in the following. method.

Authorized licensed use limited to: East Carolina University. Downloaded on June 15,2021 at 16:42:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING

TABLE VI TABLE VIII


A CCURACY S CORE OF M ODEL T ESTING W ITH N EW D ATA U SING C OMPARATIVE A NALYSIS OF THE P ROPOSED IDS W ITH O THER IDS
E MBEDDED M ETHOD AVAILABLE IN THE L ITERATURE

TABLE VII
A CCURACY S CORE OF M ODEL T ESTING W ITH N EW D ATA U SING
C ORRELATION M ETHOD

Normal Scenario: Logistic regression and RFC s always


gave the correct results. SVM gave the correct prediction
nine times, and one time, it predicted it as attack 3. DTC
also predicted the same. Naive Bayes was also off one time.
However, the polling result was always clearly predicted a
normal scenario.
Attack 1 (Hello Flood Attack): All classifiers except DTC
gave correct prediction all the ten times. The decision tree
was wrong all the ten times. However, polling gave the correct
result in the very first attempts.
Attack 2 (DIS Attack): This is the only scenario that was not
very clearly predicted in the first attempt. All the classifiers
predicted different results. No clear polling result in our IDS
leads to 20 iterations of prediction. Since the decision tree
and RFCs always gave correct prediction, the net result after
20 iterations gave a correct prediction of 2 (19 times it was
2).
Attack 3 (Increased Version Attack): Random forest, naive
Bayes, and LRCs always gave correct results. SVM and
decision tree were off the prediction, but polling always got
the correct result.
Attack 4 (Decreased Rank Attack): Random forest, decision
tree, and LRCs always gave correct results. SVM and naive
Bayes were off the prediction, but polling always got the
correct result.
b) Results of the model built using correlation method:
Table VII lists the results of the model built using the Attack 1 (Hello Flood Attack): All classifiers except DTC
correlation method. gave correct prediction all the ten times. The decision tree
Normal Scenario: GNBs always gave the correct results. was wrong all the ten times. However, polling gave the correct
SVM and logistic regression gave correct prediction nine result in the very first attempts.
times, and one time, it predicted it as attack 2. DTC was also Attack 2 (DIS Attack): This is the only scenario that was
off one time. However, the polling result was always clearly not very clearly predicted in the first attempt. All the clas-
predicted a normal scenario. sifiers predicted different results, mainly switching between

Authorized licensed use limited to: East Carolina University. Downloaded on June 15,2021 at 16:42:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SHARMA et al.: NOVEL IDS FOR RPL-BASED CPSs 7

class 2 and class 3. No clear polling result in our IDS in Data are turned into single-row features in the form of.csv
the first attempt leads to 20 iterations of prediction for the file. This.csv file is passed through the IDS trained using five
final results. Since RFCs always gave correct prediction, and different supervised machine learning classifiers, i.e., RFC,
logistic regression was predicting the correct class most of the SVM, DTC, GNB, and LRC. As the result, the IDS may detect
time, the net result after 20 iterations gave a correct prediction that the network is without any intrusion or is having one of the
of 2 (19 times it was 2). four attacks, i.e., Hello flood, DIS attack, increased version,
Attack 3 (Increased Version Attack): Random forest, deci- or decreased rank attacks. The model is also capable of detect-
sion tree, and LRCs always gave correct results. SVM and ing a new attack that does not belong to the abovementioned
naive Bayes were off the prediction, but polling always got categories.
the correct result. The high capability of detecting the known and new attacks
Attack 4 (Decreased Rank Attack): Random forest, naive is supported by experimental results. Out of the 20 instances of
Bayes, and LRCs always gave correct results. The decision the new attack, only three instances were predicted as class 3,
tree predicted the correct majority of the time. SVM was off and the rest all were classified as a new attack, i.e., giving
the prediction, but polling always got the correct result. 85% accuracy of predicting a new attack. The new attack is
not classified yet. It may be a new attack or a combination of
C. Layer 3: Model Deployment many others. Table VIII lists the comparison of the proposed
IDS with the other ones discussed in the literature.
The IDS is deployed at the node that is observing the traffic
Future work includes the plan for extending and remodeling
flow of the entire network. The data are collected every 15 min
the IDS to be able to classify the new attack from the archive
for checking an intrusion. The data in the form of a.pcap file
and name it as another class for the predictions later. This
are used and passed through the defined algorithms (1–5) to
may be done automatically based on reinforcement learning
get a brand new.csv testing data file. These testing data are
techniques.
passed through the model for the final prediction.
1) Detecting a Known Attack: The performance for the
R EFERENCES
normal traffic and the four defined attacks was similar to the
previous section. [1] K. C. Lee et al., “A comprehensive evaluation of RPL under mobility,”
Int. J. Veh. Technol., vol. 2012, pp. 1–10, Mar. 2012.
2) Detecting an Unknown Attack: We introduced a new [2] T. Clausen, U. Herberg, and M. Philipp, “A critical evaluation of the
attack in the network. We got the.pcap files for 20 different IPv6 routing protocol for low power and lossy networks (RPL),” in
simulations and created new testing set for these 20 runs. Proc. IEEE 7th Int. Conf. Wireless Mobile Comput., Netw. Commun.
(WiMob), Oct. 2011, pp. 365–372.
When passed through the model, the prediction was given as [3] O. Gaddour and A. Koubâa, “RPL in a nutshell: A survey,” Comput.
follows. Netw., vol. 56, no. 14, pp. 3163–3178, Sep. 2012.
[4] A. Raoof, A. Matrawy, and C.-H. Lung, “Routing attacks and mitigation
1) None of the models predicted class 0. methods for RPL-based Internet of Things,” IEEE Commun. Surveys
2) RFC predicted 2 or 4 several times. Tuts., vol. 21, no. 2, pp. 1582–1606, 2nd Quart., 2019.
3) SVM predicted class 1 all the time. [5] L. Wallgren, S. Raza, and T. Voigt, “Routing attacks and countermea-
sures in the RPL-based Internet of Things,” Int. J. Distrib. Sensor Netw.,
4) DTC predicted 3 or 4 a few times. vol. 9, no. 8, Aug. 2013, Art. no. 794326.
5) GNB predicted it as 3 all the time. [6] S. Raza, L. Wallgren, and T. Voigt, “SVELTE: Real-time intrusion
6) The logistic regression model classified it as class 2 all detection in the Internet of Things,” Ad Hoc Netw., vol. 11, no. 8,
pp. 2661–2674, Nov. 2013. [Online]. Available: http://www.
the time. sciencedirect.com/science/article/pii/S1570870513001005
As the final result of polling, there was an incorrect pre- [7] C. Cervantes, D. Poplade, M. Nogueira, and A. Santos, “Detection of
diction only three times, i.e., class 3 was predicted; 17 times, sinkhole attacks for supporting secure routing on 6LoWPAN for Internet
of Things,” in Proc. IFIP/IEEE Int. Symp. Integr. Netw. Manage. (IM),
the result came as a new attack leading to 85% accuracy. May 2015, pp. 606–611.
Table VIII presents the comparison of this new IDS with [8] M. N. Napiah, M. Y. I. B. Idris, R. Ramli, and I. Ahmedy,
other IDS’s reviewed in the literature. “Compression header analyzer intrusion detection system (CHA–
IDS) for 6LoWPAN communication protocol,” IEEE Access, vol. 6,
pp. 16623–16638, 2018.
[9] P. Pongle and G. Chavan, “Real time intrusion and wormhole attack
IV. C ONCLUSION detection in Internet of Things,” Int. J. Comput. Appl., vol. 121, no. 9,
The 6LoWPAN network may suffer from many intrusion pp. 1–9, Jul. 2015.
[10] I. Ahmad, M. Basheri, M. J. Iqbal, and A. Rahim, “Performance com-
threats. The novel IDS, as proposed in this article, is built parison of support vector machine, random forest, and extreme learning
using supervised machine learning techniques and is built and machine for intrusion detection,” IEEE Access, vol. 6, pp. 33789–33795,
tested on Cooja-Contiki. The IDS is successfully capable of 2018.
[11] S. M. Kasongo and Y. Sun, “A deep learning method with filter
detecting any intrusion in the network. Data from each net- based feature engineering for wireless intrusion detection system,” IEEE
work are collected after 15–20 min, using several algorithms. Access, vol. 7, pp. 38597–38607, 2019.

Authorized licensed use limited to: East Carolina University. Downloaded on June 15,2021 at 16:42:10 UTC from IEEE Xplore. Restrictions apply.

Vous aimerez peut-être aussi