Supervised Feature Selection Techniques in
Network Intrusion Detection: a Critical Review
M. Di Mauroa,<, G. Galatrob, G. Fortinocand A. Liottad
aDepartment of Information and Electrical Engineering and Applied Mathematics (DIEM), University of Salerno, 84084, Fisciano, Italy
bAmazon AWS, Belgard Retail Park, Tallaght, Dublin, Ireland
cDepartment of Informatics, Modeling, Electronics and Systems, University of Calabria, Italy
dFaculty of Computer Science, Free University of Bozen-Bolzano, Italy
ARTICLE INFO
Keywords :
Feature Selection
Machine Learning
Network Intrusion Detection
Network PerformanceABSTRACT
Machine Learning (ML) techniques are becoming an invaluable support for network intrusion de-
tection, especially in revealing anomalous flows, which often hide cyber-threats. Typically, ML al-
gorithms are exploited to classify/recognize data traffic on the basis of statistical features such as
inter-arrival times, packets length distribution, mean number of flows, etc. Dealing with the vast
diversity and number of features that typically characterize data traffic is a hard problem. This re-
sults in the following issues: i)the presence of so many features leads to lengthy training processes
(particularly when features are highly correlated), while prediction accuracy does not proportionally
improve; ii)some of the features may introduce bias during the classification process, particularly
those that have scarce relation with the data traffic to be classified. To this end, by reducing the fea-
ture space and retaining only the most significant features, Feature Selection (FS) becomes a crucial
pre-processing step in network management and, specifically, for the purposes of network intrusion
detection. In this review paper, we complement other surveys in multiple ways: i)evaluating more
recentdatasets(updatedw.r.t. obsoleteKDD 99)bymeansofadesigned-from-scratchPython-based
procedure; ii)providingasynopsisofmostcreditedFSapproachesinthefieldofintrusiondetection,
includingMulti-ObjectiveEvolutionarytechniques; iii)assessingvariousexperimentalanalysessuch
as feature correlation, time complexity, and performance. Our comparisons offer useful guidelines
to network/security managers who are considering the incorporation of ML concepts into network
intrusion detection, where trade-offs between performance and resource consumption are crucial.
1. Introduction
With the rapid growth of digital technology and com-
munications, we are overwhelmed by network data traffic,
whicharediverseformediatype(e.g. video,voice,text,sen-
sory,etc.),andoriginatefrom(andaretransportedthrough)
abroadrangeofsources(e.g. mobilenetworks,cloudinfras-
tructures,InternetofThings,etc.). Consequently,wehandle
high-dimensionality data, calling for increasingly more so-
phisticated classification methods [1, 2].
Typically,werefertohighdimensionalitywhenwedeal
with data whereby a large number of features may be ex-
tracted, to the point that the features may even exceed the
numberofobservations. Thisleadstomajorissues,particu-
larly the massive increase in training times.
To this end, Feature Selection (FS) is a promising re-
searchdirection,lookingatwaystoreducethefeaturespace
in order to pinpoint only the most significant features. As
a fundamental pre-processing step in machine learning, FS
is gaining prominence in network management and, specif-
ically, for the purposes of network intrusion detection and
network traffic classification problems [3, 4, 5, 6].
Moregenerally,FSfindsanevenmuchbroaderapplica-
bilityinfieldasdiverseasbioinformatics[7,8,9,10],image
recognition/retrieval [11, 12, 13, 14, 15, 16, 17], fault diag-
<Corresponding author
mdimauro@unisa.it (M. Di Mauro); galatrogiovanni@gmail.com (G.
Galatro); g.fortino@unical.it (G. Fortino); antonio.liotta@unibz.it (A.
Liotta)nosis[18,19],textmining[20,21,22]and,interestingly, in
network traffic analysis/classification, whose pertinent bib-
liography will be covered more closely in the following.
Machinelearningenginescanbeeasilyembeddedinnet-
work intrusion detection systems (IDS), which represent an
essential part of network infrastructures to guarantee secu-
rity [23, 24, 25] and availability [26, 27, 28, 29, 30, 31, 32,
33, 34]. Specifically, modern NIDS can be equipped with
softwareprobesinchargeofanalyzingnetworktrafficonthe
basis of some characterizing features such as: distribution
of inter-arrival times, distribution of packet sizes, presence
of specific TCP/IP flags, percentage of forward/backward
flows. Suchstatisticalinformationisinstrumentaltoreveal-
inganomaloustraffic,whichisoftenbehindDistributedDe-
nial of Service (DDoS) attacks [35, 36], covert Voice-over-
IPsessions[37],threatdiffusion[38]andPeer-2-Peertraffic
[39, 40, 41]. In many cases, these flows would pass unob-
servedthroughotherconventionalsignature-basedanalyses.
In principle, a larger number of features would allow to
perform more granular analyses; yet, two main drawbacks
emerge. First, a proliferation of correlated features leads to
levels of redundancy that are useless, while resulting in in-
creasingly longer training times. Also, not all features are
valuable in characterizing traffic, incurring bias during the
classification step.
Ifweconsider,asanexample,somenovelprobessuchas
ISCXFlowMeter[42],thesecanactuallygeneratemorethan
80features to characterize data traffic, which makes it ex-
Di Mauro et al.: Preprint submitted to Elsevier Page 1 of 19arXiv:2104.04958v1  [cs.CR]  11 Apr 2021Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
tremelydifficultforthenetworkanalysttodealwiththisvast
amount of information. This is where an FS pre-processing
step becomes invaluable.
Across the scientific literature, many works are devoted
to surveying machine learning algorithms with applications
tointernettrafficclassification. Yet,fewattemptshavebeen
made in the field of FS techniques, where two main short-
comings emerge: i)many works analyze and compare FS
algorithms applied to non-specific datasets [43, 44, 45]; ii)
areeitherbasedonoutdateddatasets,orlacktheexperimen-
taldimension. Overall,notmuchnovelalgorithmshavebeen
compared(refertodetailsprovidedinTable 1,discussedfur-
ther ahead). Aimed at filling these gaps, our paper provides
the main following contributions:
•We perform an experimental analysis of more re-
centdatasets(classicliteraturefocusesonthe 20-year
old KDD 99dataset) by means of our designed-from-
scratchPythonbasedroutine. Thisallowsusto i)per-
form cleaning, re-balancing, and data mixing opera-
tions,and ii)automaticallyinteractwithspecificML-
based engines.
•We present a variety of experimental results (in-
cluding: feature correlation, time complexity, perfor-
mance) aimed at critically comparing selected fam-
ilies of FS algorithms, ranging from classic ones
(rank search, linear forwarding selection) to modern
bio-inspired algorithms (genetic search, ant colonies,
multi-objectiveevolutionary),thusgoingwellbeyond
a conventional literature survey.
Our experimental-based, comparative assessment finds
fruitfulapplicationsinthefieldofnetwork/securitymanage-
ment, where machine learning techniques are proved to of-
feraprecioussupporttointrusiondetectiontasks,andwhere
criticaltrade-offsbetweenperformanceclassificationandre-
source consumption arise.
Thepaperisorganizedasfollows: Section 2offersagen-
eralperspectiveonFSmethods,inlinewithotheraccredited
taxonomies. In Section 3we analyze how diverse FS tech-
niqueshavebeenappliedintheliterature,oftenbasedonthe
oldKDD 99dataset. Section 4proposesan excursusofpop-
ular FS algorithms (classic and modern) along with some
necessary technical details. In Section 5we analyze the ex-
ploitednoveldatasetsbygroupingthemostrelevantfeatures
in families. In Section 6we perform an experimental-based
comparative analysis of prominent FS algorithms, provid-
ing performance, feature correlation, and time-complexity
figures. Finally, Section 7draws conclusions and provides
future-direction indications.
2. Overview of Feature Selection
Feature Selection refers to that set of techniques and
strategies that allow to optimize the feature space, namely,
ann-dimensionalspacewhereeachsampleisrepresentedas
a point. When dealing with large feature spaces, the anal-
ysis of data, which starts from their representative features,can be tremendously time and resource consuming. Hence,
the need to devise suitable FS strategies aimed at eliminat-
ingi/irrelevant features, namely, those features that are not
actually needed to build an optimal feature subset; and ii/
redundant features, namely, those features that strongly de-
pend on other features [46].
Feature selection can be considered as a special case of
feature extraction methods [47]. The latter refer to a set of
techniques (e.g. PCA, Single Value Decomposition, Linear
Discriminant Analysis and others) useful to transform the
original feature space in a new one aimed at alleviating the
effectsofthenotorious curseofdimensionality problem[48,
49]. Itisimportanttonotethatfeatureextractioncomeswith
thecriticalriskofobtainingatransformedfeaturespacethat
couldloseitsoriginalphysicalmeaning,whereasclassicFS
aims at preserving it [50, 51].
A more formal definition of the FS problem follows.
Given a feature set X= ^xi:i= 1;§;N`, find a subset
SK= ^xi1;xi2;§;xiK`withK < N, where an objective
function Y() is optimized, namely:
^xi1;xi2;§;xiK` = arg max
K;ik[Y^xi:i= 1;§;N`]:(1)
Unlessotherwisestated,theproblemistoselecttheoptimal
subsetoffeatures(accordingtoaspecificcriterion)fromthe
initialset,wheretwostepsaretypicallyperformed: thefirst
oneinvolvesasearchstrategytopinpointcandidatesubsets;
thesecondoneinvolvesanobjectivefunctiontoevaluatethe
selected candidate subsets. The latter can be split in two
types [52, 53, 54, 55]: i/filters, referring to objective func-
tions that rely on properties of the data by evaluating the
information content (e.g. correlation measures, inter-class
distance);ii/wrappers, referring to objective functions that
exploittrainingmodelsbystartingfromasubsetoffeatures
and,then,addingorremovingfeaturesbasedontheprevious
model.
A commonly accepted taxonomy of FS methods is pre-
sented in [56], where three classic approaches have been
identifiedas supervised ,unsupervised ,andsemi-supervised .
According to the supervised approach, labeled data are ex-
ploitedtosingleoutafeaturesubset,consideringspecificcri-
teriaformeasuringthefeaturesimportance. Conversely,un-
supervisedtechniquesseektounveiltheintrinsicdatastruc-
turetoselectthemostsignificantfeatures,withoutassuming
anya prioriknowledge [57].
Finally, the semi-supervised approach is based on a
mixedstrategy,strivingtoenrichanunlabeledsetwithsome
labeled data, so as to improve the FS phase. Both the un-
supervised and the semi-supervised approaches exhibit the
drawback of neglecting potential correlations among fea-
tures, resulting in the analysis of sub-optimal sets. This
may prove critical when dealing with traffic analysis, where
we need to take into account statistical-based features (e.g.
inter-arrivaltimesvariance,averagepacketlength,etc.) and
deterministic ones (e.g. IP addresses, port numbers, etc.).
Let us consider, for example, a particular kind of traffic di-
rected towards a fixed destination port for a certain period
Di Mauro et al.: Preprint submitted to Elsevier Page 2 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
Table 1
Prominent related work surveying FS techniques applied to Network Intrusion Detection.
Authors Experiments Single/Multi Class Description
Wang et al. [58] Performance analysis Single Class Feature Selection performed on the KDD99
dataset, by applying C4.5 and Bayesian Net-
work algorithms
Janarthanan et al.
[59]Performance analysis Single/Multi Class Empirical selection of features by performing
tests on KDD99 and UNSW-NB15 datasets
by means of Random Forest algorithm
El-Khatib [60] Performance analysis Single/Multi Class Feature selection combining Filter/Wrapper
methods for Wireless IDS. Tests have been
performed on a WLAN environment
Chen et al. [61] Performance analysis, Time
analysisSingle Class Correlation-based Feature selection using the
KDD99 dataset
Nisioti et al. [62] N/A N/A Classic survey on ML-based techniques (in-
cludingFS)withpointerstodetailedsources,
but with no experiments
Iglesias et al. [63] Performance analysis Single/Multi Class Comparison among 4FS algorithms on the
NSL-KDD dataset
Singh et al. [64] Performance analysis, Time
analysisSingle Class Comparison among various FS algorithms on
the KDD99 dataset
Bahrololum et al.
[65]Performance analysis Single/Multi Class Comparison among PSO, Decision Tree,
FlexibleneuraltreealgorithmsontheKDD99
dataset
Dhote et al. [66] N/A N/A Limited survey of FS techniques applied to
network traffic, with pointers to detailed
sources, but with no experiments
This work Performanceanalysis, Feature
Correlation analysis, Time
analysisSingle/Multi Class Feature Selection performed on the CIC-IDS-
2017/2018dataset, byconsidering 9FSalgo-
rithms from classic (Rank, Scatter) to mod-
ern (Genetic, Multi-Objective Evolutionary)
oftime. Anunsupervisedapproachcouldleadtoasubsetof
features which does not include the destination port. Thus,
acrucial(aswellasdeterministic)pieceofinformationgets
lost.
Ontheotherhand,asupervisedapproachcanofferopti-
malresults,providedthatthedataarecorrectlylabeled. This
case typically occurs in a controlled network environment,
where, with the help of network analyzers, it is possible to
automatically label the type of passing data traffic.
Because of the great potential of supervised methods,
andthankstotheavailabilityofsuitablelabeleddatasets,we
have decided to focus our experimental-based comparative
evaluation on supervised FS methods.
3. Related Work on Feature Selection applied
to ML-based Intrusion Detection
Most scientific literature involving ML approaches is
typically focused on the proposal of algorithms or tech-
niquesforclassification/detectionofspecificnetworktraffic
([67,68,69,70,71,72,73,74]). Unfortunately,thestraight
applicationofmachinelearningtonetworktrafficanalysisis
not always feasible, due to the broad variety of traffic types,
which leads to unmanageable feature spaces. Intuitively, in
fact, a highly diversified traffic (multimedia, asynchronous,bursty, etc.) requires a large set of features able to capture
the variegated “nature" of the different flows. That is why
appropriatepre-processingsteps(i.e. FS)playacrucialrole,
thus a significant portion of ML-based literature has shown
interest in FS techniques.
Forinstance,toimprovetheperformanceofIDSframe-
works, the authors of [75] propose a mixed strategy in-
volvingPrincipalComponentAnalysisandfuzzyclustering
withKNN-basedFStechniques. Acorrelation-basedFSap-
proachcoupledwithaSupportVectorMachine(SVM)clas-
sifierisproposedin[76]tobuildacloud-basedIDS.AnIDS
based on deep learning methods, along with a filter-based
FS algorithm, is introduced in [77]. Similarly, a Convolu-
tional Neural Network (CNN) approach is exploited in [78]
to select traffic features from raw data sets, improving the
accuracy of an intrusion detector. Yu and Liu propose a
mutualinformation-basedalgorithmthatcananalyticallyse-
lect optimal features for classification, by handling linearly
andnon-linearlydependentdata[79]. Again,FSisexploited
jointlywithArtificialNeuralNetworks[80]andDeepNeural
Networks [81], respectively, to improve IDS performance.
Furthermore,authors in[82]embedin anIDStwoFSalgo-
rithms that are compared against mutual information-based
methods.
A major shortcoming of these works is that methods
Di Mauro et al.: Preprint submitted to Elsevier Page 3 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
are validated and compared through the outdated KDD 99
dataset. This contains information about older network at-
tacks (that have been mitigated by now) or, in some cases,
adopts an updated version of KDD 99, namely NSL-KDD
[83]. Yet, although NSL-KDD adds some improvements
onto KDD 99(e.g. no redundant records, better balancing
between training and test set, etc.), it does not take into ac-
count features that characterize novel cyber attacks. Other
works in the field of machine learning applied to intrusion
detectionrelyonmorerecentdatasetssuchasUNSW-NB15
[84, 85, 86, 87]. Although quite recent, such a dataset has
two limitations: first, the traffic has been collected in a re-
duced testbed; and secondly, the number of features is lim-
ited to 49, which is too small to appreciate the effectiveness
of feature selection techniques. In contrast, in our work we
relyonanup-to-datedatasetfromtheCanadianInstitutefor
Cybersecurity [88] which has the following benefits: i/it
containsdatatrafficgatheredoveravastnetworkarea;and ii/
itaccountsforabout 80features,allowingtoextensivelytest
the feature selection algorithms. Additional details about
this dataset are provided in Section 5.
Going more specifically into the set of works that share
withthispapertheaimtocompareorsurveyFSmethodsfor
intrusiondetection,wehavecollectedthesignificantpapers
in Table 1. In it, we have identified the material covered in
the literature, which helps appreciating the contributions of
our paper. The first column of Table 1points to the source;
thesecondcolumnhighlightsthetypeofexperimentalanal-
ysis (if any); the third column pinpoints the type of datasets
utilized (single or multi-class); the last column provides a
concise description of the surveyed material.
4. Review of Feature Selection Algorithms
under Scrutiny
In this section, we briefly describe the algorithms under
scrutiny,whichbelongtodifferentfamiliesofFStechniques,
rangingfromclassicrank-guided(Rank,LinearForwardSe-
lection) and meta-heuristic (Tabu, Scatter, Particle Swarm)
ones, to nature-inspired algorithms (Ant, Cuckoo), and up
to modern techniques (Genetic, Multi-Objective Evolution-
ary).
For each algorithm, we provide a brief recap along with
itspertinentapplicationinnetworktrafficanalysisandsecu-
rity in literature.
4.1. Rank-based Feature Selection
Algorithms belonging to this family follow an approach
based on two macro-steps: in the first one, the features are
rankedaccordingtoacertainstatisticalmeasure,whereasin
thesecondstepthealgorithmchoosesthetoprankedfeatures
(eventually partitioned in clusters). We investigate and put
to test two representative algorithms of this family: Rank
Search and Linear Forward Selection.
4.1.1. Rank Search
TheRankSearchtechniquerefersnotonlytoasingleal-
gorithm,but,toanumbrellaofmethodsabletoproducealistof attributes ranked with some criteria. One of the most ap-
plied rank-based techniques in FS relies on the Information
Gain (IG) concept [89]. Given an attribute Aand a class C,
theentropyof the class without and with prior observation
of the attribute are, respectively:
H.C/ = *É
cËCp.c/log2p.c/; (2)
and
H.CðA/ = *É
aËAp.a/É
cËCp.cða/log2p.cða/:(3)
TheIGisgivenby H.C/ *H.CðAi/,andrepresentsthe
amount of the class entropy decreasing due to the a priori
knowledge introduced by i-th attribute.
Over network traffic analysis literature, the rank search
methodhasbeenwidelyexploitedinconjunctionwithstan-
dard machine learning algorithms typically used during the
classificationstep. In[90],havingtheNSL-KDDdatasetas
input, the proposed algorithm i/evaluates the information
gain value,ii/applies fuzzy rules to remove unwanted fea-
tures,iii/calculatesthemeanvalueofIGacrossthenewsub-
setoffeatures, iv/refinesthefeaturesubsetsbyapplyingan
algorithm based on conditional probability evaluation. Au-
thors in [91] propose a detection model able to cope with
network attacks based on the feature IG ranking; once ob-
tained a satisfying feature set, a triangle area based KNN,
combiningbothSVMandgreedytechniques,isexploitedto
single out even more discriminative and useful features. A
combination between an IG-based feature selection method
andaC4.5-basedclassifierisadvancedin[92],wherethere-
sultingalgorithmhasbeenoptimized(intermsofpowercon-
sumption) to reveal DoS attacks in ad hoc networks. Simi-
larensemblehasbeenexploitedin[93],whereauthorscon-
siderabroadersetofclassifiers(C4.5,RandomForest,Naive
Bayes, etc.) and where the analysis focuses on generic mal-
ware detection. Again, a mixed Genetic/KNN approach for
intrusion detection presented in [94] benefits from a feature
reductionprocedureobtainedapplyingID3algorithmtofind
higher IG.
4.1.2. Linear Forward Selection
Suchatechniquetoreducethefeaturespacedimension-
ality has been presented in [95]. Linear Forward Selection
(LFS) can be considered as an improvement of Sequential
ForwardSelection(SFS)methodwhichstartswithanempty
set of features and sequentially adds one feature at a time.
In order to face the O.N2/complexity of SFS, in LFS the
number of considered features at each step does not exceed
a certain user-specified constant, thus, the resulting perfor-
mance is impressively ameliorated. Two methods for limit-
ing the number of features have been implemented in LFS
algorithm:i/FixedSet,whereascoreobtainedbymeansof
a wrapper evaluator is used to choose the top- kranked at-
tributes, thus, the maximum number of evaluations reduces
tok_2.k+ 1/;ii/Fixed Width , where, at each forward
Di Mauro et al.: Preprint submitted to Elsevier Page 4 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
selection step, the number of features is increased by one,
suchthatthesetofcandidateexpansionsconsistsofthebest
kfeatures not been selected so far during the search. In
such case, the maximum number of evaluations amounts to
Nk*k_2.k+1/. ForwardSelection-basedtechniqueshave
beenexploitedacrossthetrafficclassificationdomainforde-
tection of malicious application on Android [96], detection
ofzero dayattacks [97], designing novel anomaly detection
systems [98].
4.2. Meta-heuristic Feature Selection
Techniquesbelongingtothisfamilyrelyontheprinciple
that it is possible to select heuristics , namely, approximate
algorithms searching for a sufficiently good solution to an
optimization problem, useful when some constraints arise
(e.g. limited computational resources, incomplete informa-
tion). Weselectthreerepresentativealgorithmsofthisfam-
ily: Tabu Search, Scatter Search, and Particle Swarm Opti-
mization.
4.2.1. Tabu Search
Tabu Search (TS) algorithm has been originally pro-
posed in [99], and, then, it has been extended and exploited
to solve practical optimization problems [100, 101, 102].
TS is based on a metaheuristic method able to pilot a lo-
cal heuristic search in exploring the space of solutions be-
yond the local optimality. Two main features characterize
TS:adaptive memory andresponsive exploration . The for-
mer allows to perform local choices guided by information
collected during the search, whereas, the latter allows to
make strategic choices. In other words, TS exploits a local
search procedure combined with memory-based strategies,
thus,theissueofgettingtrappedinlocaloptimalsolutionsis
avoided. To implement the memory-based mechanism, TS
builds a map of recently visited solutions called Tabu List
(TL). A simplified TS algorithm is illustrated below:
Algorithm 1: Tabu Search
1. Given a function f.x/to be optimized over a set
X, start from initial solution x0ËX, initialize
Tabu List (TL), and initialize a counter i= 0.
2. GivenN.xi/ÏXtheneighborhood ofxi,N.xi/
can be reached by xiby means of a move
operation. Thus, generate a neighborhood move
listM.Xi).
3. Givenxi+1the best solution in M.Xi), update
TL.
4. In case stopping conditions are met, terminate.
Otherwise, repeat Step 2.
It is interesting to notice that, the memory structures char-
acterizing TS method operate according four dimensions
[100]:
•Recency: concerns the ability of keeping track of so-
lutions attributes that have changed across the recent
past.•Frequency : involves the mechanism exploited to
broad the foundation for selecting preferred moves.
•Quality: pertainstothecapacityofdiscriminatingthe
meritofsolutionsvisitedduringthesearchoperation.
•Influence: takes into account the impact of choices
performed during the search.
Inthefieldofnetworktrafficclassification,TShasbeen
exploitedin[103]jointlywithfuzzytechniquesaimedatop-
timizing the exploration of feature search space in intrusion
detection problems. A combination of TS technique and
KNNispresentedin[104],whereKNNisinitiallyexploited
to generate a subset of non-redundant features, whereas, TS
isusedtorefinetheobtainedsubset. Again,authorsin[105]
propose a wrapperFS algorithm, dubbed GATS-C4.5, that
embedsanhybridGeneticandTabu-basedmethodasfeature
selection strategy and a supervised ML algorithm (C4.5) as
the evaluation function. Similar combinations between TS
andGenetictechniqueshavebeenusedin[106]wheresome
tests (vs pure Genetic algorithms) across DARPA dataset
have been carried out, and in [107] where SVM has been
adopted as a classification criterion.
4.2.2. Scatter Search
Scatter Search is a metaheuristic algorithm involving
memory-based mechanisms similar to those exploited in
Tabu Search [108]. The main strategy relies on an iterative
process that organizes high-quality optimal solutions into
subsets, and where five “methods" emerge: i/Diversifica-
tion,tocreateasetofdifferenttrialsolutionsbyexploitinga
seed solution; ii/Improvement , to convert a trial solution in
one or more improved solutions; iii/Reference Set Update ,
to create and keep update a reference set of best solutions;
iv/SubsetGeneration ,tomanipulatethereferencesetaimed
at deriving a subset of solutions; v/Solution Combination ,
to linearly re-combine the solutions on the basis of subsets
obtained with the previous methods.
As an effective FS procedure, Scatter Search has been
exploitedtogetherwithNLPsolversforglobaloptimization
[109],withroughsetsfortheimplementationofcreditscor-
ing mechanisms [110], or in a parallelized fashion to im-
prove the feature subset selection problem [111]. Again,
ScatterSearchhasbeenprofitablyexploitedinissuesinvolv-
ingcreditcardsfrauddetection[112],andsoftwaresecurity
characterization [113].
4.2.3. PSO Search
The Particle Swarm Optimization (PSO) algorithm has
beenoriginallydiscoveredin[114],throughsimulationscar-
riedoutacrossasimplifiedsocialmodelaimedatreproduc-
ing the behavior of birds flocking.
Then,thepopulationofagentsbecamemoresimilartoa
swarm than flock, and single individuals were named parti-
clesthat represent the candidate solutions. In a mathemat-
ical form, given AÏRnthe search space, and f:A, , ™
Y Ó R, the swarm is defined by a set S= ^x1;x2;§;xN`
ofNparticles, where xi= .xi1;xi2;§;xin/TËAwith
Di Mauro et al.: Preprint submitted to Elsevier Page 5 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
i= 1;2;§;N. It is assumed that particles are able to it-
eratively move within search space Aby means of a ve-
locity parameter defined as vi= .vi1;vi2;§;vin/Twith
i= 1;2;§;N. In PSO method, it is also defined a mem-
ory set P= ^p1;p2;§;pN`containing the best positions
pi= .pi1;pi2;§;pin/T(withi= 1;2;§;N) visited by
each particle. Given tthe time counter, the current position
and velocity for particle iare, respectively, xi.t/andvi.t/,
whereaspi.t/ = arg mintfi.t/. Accordingly, PSO is defined
by the following equations:
vij.t+ 1/ =vij.t/ +
1r1[pij.t/ *xij.t/](4)
+
2r2[g.t/ *xij.t/];
xij.t+ 1/ =xij.t/ +vij.t+ 1/; (5)
where:r1andr2denote random variables uniformly dis-
tributed in [0,1], 
1is thecognitive parameter which affects
thestepsizethattheparticletakestowardsitsbestcandidate
solutionpij.t/, and
2is thesocialparameter which affects
thestepsizethattheparticletakestowardstheswarm’sbest
solutiong.t/. At each iteration, best positions pi.t+ 1/are
updated as well, namely
pi.t+1/ =h
n
l
njxi.t+ 1/iff.xi.t+ 1//ff.pi.t//;
pi.t/otherwise.(6)
InthefieldofFSappliedtonetworktrafficanalysis,PSO
hasbeenprofitablyexploitedjointlywithclassificationtech-
niques. It is the case of [115, 116], where a particle swarm
selectionmethodhasbeenexploitedwithSVM-basedclassi-
fiers. Again,ahybridFSmodelbasedonPSOandRandom
Foresthasbeenexploitedin[117],whereindependentmea-
suresandalearningalgorithmareexploitedtoevaluatefea-
ture subsets. Interesting is also an evolution of classic PSO
advanced in [118], and dubbed Accelerated PSO, amenable
to deal with FS on Big Data streams.
4.3. Nature-inspired Feature Selection
Although relying on meta-heuristic concepts, this fam-
ily of algorithms takes inspiration from the nature ecosys-
tem, where many animal species exhibit impressive behav-
iorsaimedatoptimizingtheirlifecycle. WeassessAntOp-
timization and Cuckoo search as representative algorithms
of such family.
4.3.1. Ant Search
This technique, originally proposed in [119], is inspired
byantscoloniesbehavior,wheretheoptimizationproblemis
solvedbya“colony"ofcooperatingagents. Analysescarried
out by ethologists showed that, following pheromone trails,
eachantisabletofollowaprecedingantwhichreleasessuch
substance, thus, a whole ant colony is able to self-organize
itself. The emerging collective behavior relies on a positive
feedback loop: the probability which an ant chooses a cer-
tainpathincreaseswiththenumberofantsthatchoosingthesamepath,sincethetrailiscontinuouslyreinforcedwithnew
pheromone. The Ant algorithm can be profitably exploited
in feature selection problems by evaluating the probability
pk
ijthat thek-th “ant" could arrive to feature jby starting
from feature i, namely,
pk
ij=h
n
l
nj[ij][ij]
³
kËUk[ik][ik]ifjËUk;
0otherwise,(7)
where,Ukisthesetoffeasibleattributes(notvisitedyet), ij
istheamountofpheromoneacrossthe ijpath,ijrepresents
theheuristicinformationfortheselectedattribute j,and
areparametersinchargeofcontrollingpheromonetrialsand
heuristic information, respectively.
As regards the FS problem in network traffic analysis,
Ant-based methods have been used in: [120] where a SVM
classifier is adopted on KDD99 dataset, after a feature re-
duction obtained through ACO (Ant Colony Optimization)
technique; in [121] where an ACO-based feature selection
method allows to deal with big streamed data; [122] where
animprovedACO-basedalgorithm(namedFACO)hasbeen
designed and tested across the classic KDD99 dataset.
4.3.2. Cuckoo Search
This technique is inspired to the brood parasitism strat-
egycharacterizingsomecuckoosspecies. Inparticular,such
specieslaytheireggsinthenestsofotherbirds(hosts). Since
hostbirdscanengageaconflictastheyrecognizealieneggs,
particularcuckoospecieshaveevolvedinsuchawaythatfe-
malesareableinmimickingcolourandsizeofeggsofsome
host species, thus the hosts are cheated, and the probabil-
ity of cuckoos reproductivity grows. Considering an egg in
a nest as a solution, three idealized rules emerge in Cuckoo
Searchprocedure[123]: i/eachcuckoolaysoneeggattime,
in a randomly chosen nest; ii/the bests nests (having high-
quality eggs) are candidate to carry over next generations;
iii/thenumberofnestsisfixedand,asahostbirddiscovers
alien(cuckoo)eggswithaprobability pd,itgetsridofit. The
aimofthealgorithmistoexploitnew(andeventuallybetter)
solutions in place of not-so-good solutions in the nest. New
solutions xt+1
ifor cuckooiare obtained through the follow-
ing expression which represents the stochastic equation for
random walk, namely,
xt+1
i=xt
i+äL./; (8)
where, > 0is a scale factor, the product ärefers to en-
trywise multiplications, whereas Lis the Lévy distribution
with ( 1<f3).
Cuckoo search method has been exploited in network
trafficanalysispairedwithvarioustechniquesandtechnolo-
gies. In [124], authors propose an algorithm that uses PCA
and Cuckoo Search to reduce the feature space and to op-
timize the clustering center selection. A Cuckoo-based FS
algorithm is proposed in [125] to preprocess network data
Di Mauro et al.: Preprint submitted to Elsevier Page 6 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
aimedatimprovingtheIDSdetectionaccuracyinclouden-
vironments. A Cuckoo search strategy has been also used
in [126] to optimize Artificial Neural Networks when deal-
ing with traffic anomaly detection issues. Again, coupled
with SVM, Cuckoo search has been adopted in FS to deal
with problem of phishing mail detection [127]. Recently,
extended versions of Cuckoo Search algorithm have been
advanced to cope with classification of tweetsin sentiment
analysis[128],ortodefeatattacksinSoftwareDefinedNet-
work infrastructures [129].
4.4. Evolutionary Feature Selection
Suchfamilyofalgorithmsisinspiredbynaturalselection
theory, claiming that living organisms survived across mil-
lions of years thanks to an adaptation process. In a similar
way,thisaptitudecanbetranslatedinsearchforoptimalso-
lutions to a problem. Two exemplary tested algorithms are:
Genetic search and Multi Objective Evolutionary search.
4.4.1. Genetic Search
Genetic Algorithms (GAs) have been designed around
themid-1950s,whenbiologistsstartedtoperformcomputer-
based simulations aimed at analyzing more in deep the
evolution of genetic processes [130]. Then, GAs have
been extended to face problems ranging from neural net-
works weight estimation [131] to inequalities-based prob-
lems[132]. Apioneeringworkinthisfieldhasbeencarried
outbyHolland[133,134],and,today,manyvariantsofGAs
exist [135] and are applied in economy, computer science,
sociology.
The basic skeleton of a GA includes three operators
[136]:Reproduction, Crossover andMutation.
Reproduction refers to a process in charge of evaluating
theabilityofanindividualtobeselected(amongothers)for
reproduction, on the basis of a fitnessscore.
Crossover concerns the capability of a genetic operator
in recombining information to create new offspring. Typi-
cally, offspring is generated by exchanging genes of parents
until acrossover point is reached.
Mutation pertains to the probability that some offspring
genes could be modified or altered.
Genetic-based feature selection in network traffic analy-
sishasbeenusedinconjunctionwithmanyML-basedmeth-
ods. Authors in [137] exploit a GA-based FS approach to
optimize network traffic data before applying an artificial
neuralnetworktoperformattacksdetectionacrosscloudin-
frastructures. A combination of a genetic FS method and a
supervised classifier based on J48 algorithm is proposed in
[138]. More frequent across the scientific literature is the
couplingbetweengeneticFSandSVMclassifiersappliedto
networktrafficclassificationproblems(see[139,140,141]).
When dealing with FS problems, GAs allow to explore
the solution space by selecting the most promising regions,
thus,avoidingacostlyexhaustivesearch. Inourdomain,the
initial population is represented by the whole feature space
and the fitness function relies on the correlation among fea-tures and expressed by means of a meritindicator defined
further ahead in eq. (12).
Once entered the cycle represented in Fig. 1, the algo-
rithm calculates the fitness of each candidate solution per
iteration, selects individuals to reproduce, and generates a
newpopulationbytakingintoaccountcrossover(featurere-
combination with a certain probability), and mutation (one
featurecanbeturnedintoanotherfeaturewithacertainprob-
ability).
!"#$#%&'()*&%$#(" 
!"#"$%&"'%#'(#(&(%)
*+*,)%&(+#+-'(#.(/(.,%)0 +,%&*%$#(" 
1/%),%&"&2"'-(&#"00'
-,#3&(+#4+567'-,#3&78 -#$".//01%&*./ 
900(:#&2"'%**$+*$(%&"'
-(&#"00'/%),"0 
2.&.3$#(" 
;(#:)"'+,&'&2"'(#.(/(.,%)0 
-+$'$"*$+.,3&(+# 4.)5(6*3$#(" 
!"#"$%&"'#"<'(#.(/(.,%)0 
43$+00+/"$='>,&%&(+#8'
Figure 1: Genetic Algorithms life cycle.
4.4.2. Multi-Objective Evolutionary Search
The family of solutions concerning a multiobjective op-
timization problem (MO) includes all the elements of the
search space whose objective vectors cannot be simultane-
ously improved (Pareto optimality concept) [142]. The set
of such objective vectors is said non-dominated.
Moreformally,aMOproblemcanbeformulatedasfol-
lows: given a vector of nobjective functions fof a vector
variable xin a domain Ddefined as
f.x/ = .f1.x/;f2.x/;§;fn.x/; (9)
a decision vector xhËDis Pareto-optimal iffthere is no
xkËDsuch that:
h
n
n
l
n
njÅiË ^1;§;n`;kifhi
á
ÇiË ^1;§;n` :ki<hi:(10)
On the other hand, Evolutionary Algortihms (EAs)
can be profitably exploited in MO-based problems since
many “individuals" can search in parallel for multiple so-
lutions, with the possibility of taking advantage of similari-
tiesamongsolutionsbelongingtothesamefamily. Possible
implementationsofMO-EAtechniquesareENORA(Evolu-
tionary NOn-dominated Radial slots based Algorithm), and
NSGA (Non-dominated Sorted Genetic Algorithm) com-
pared in [143].
Applied to the FS problem, the purpose of a multi-
objective search algorithm is to discover a subset of fea-
tures (a family of solutions) being a good approximation of
the Pareto front. The MO-EA approach has been exploited
Di Mauro et al.: Preprint submitted to Elsevier Page 7 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
in network traffic classification jointly with ensemble ML-
based methods [144], where some objectives such as max-
imizing true positive rate, maximizing classification accu-
racy,minimizingfeaturenumber,andminimizingfalsepos-
itiverate,aresatisfiedsimultaneouslyandwithnoconflicts.
An NSGA-based approach to ameliorate the performance
classificationofIDSplatformshasbeenadoptedin[145]and
in[146]. Again,aMO-EAtechniquejointlywithfuzzyclas-
sifiers for coping with the traffic classification problem has
been introduced in [147].
5. The considered Datasets
One of the great issues when dealing with supervised
FSapproachesintrafficanalysis,isfindingtrainingsetsthat
are both recent and labeled. As remarked in Sect. 3, many
works rely on the obsolete KDD 99dataset [148]. Created
about 20years ago, KDD 99has been broadly employed to
validate machine learning algorithms, particularly to differ-
entiatemaliciousdatatrafficfrombenignone. However,the
KDD 99dataset does no longer reflect the characteristics of
moderndatatraffic,whichhassensiblychangedacrosstime.
Thisisverymuchthecaseofmultimediatraffic(e.g. voice,
video), being the principle revenue-making stream for ser-
vice providers but also the vehicle of covert malicious data.
Percontra ,inthisworkweconsidermorerecentdatasetsob-
tainedbymeansofCICFlowMeter[42,149],anopen-source
enginethatcanbothgatherandlabelnetworktrafficinacon-
trolled environment. Each dataset contains records labeled
eitherasBenignorMalicious(malicioustrafficisoftensplit
insub-labeledtrafficrepresentingdifferentkindsofattacks).
Specifically, we consider the following datasets:
•DDoSwhichcontainstrafficrelatingtodistributedde-
nialofserviceattacks,aimedatsaturatingthenetwork
resources of specific targets;
•Portscan which contains traffic relating to Portscan
attacks, aimed at discovering open, network device
ports;
•Webattack , including malicious traffic which imple-
mentsvariousweb-basedattackssuchasBruteForce,
Cross-Site Scripting, and Sql Injection;
•TORwhichincludestrafficpassingovertheTORnet-
work,ananonymousandprivatedatacircuitoftenex-
ploited to carry dangerous information or malicious
encrypted traffic;
•Androidwhich embeds various families of Android-
based threats (adwares, ransomwares, etc.).
We want to notice that the first four datasets are exploited
into the single class analysis, whereas the Androiddataset
is exploited in the forthcoming multi-class analysis. In this
latter analysis, we build a new dataset called MultiAndroid
(see Sect. 6:2), obtained by selecting the most relevant mo-
bile threats from the Androiddataset mixed with some be-
nign traffic. In order to avoid the issue of bias that typi-
cally arises during classification in imbalanced datasets, wehave first re-arranged the datasets. We have achieved well-
balanced benign/malicious features, spanning across about
50kinstances for each dataset. Each one contains up to 78
features, except for the TOR dataset including 30 features.
Forthesakeofconvenience,wefounditusefultogroupthe
featuresin 5macro-classes(thecompletelistoffeaturescan
be found in [150]):
•Time-based features: Forward/Backward inter-
arrivaltimes(IAT)betweentwoflows,durationofac-
tive flow (min, max, mean, std), duration of idle flow
(min, max, mean, std), etc.;
•Byte-based features: Forward/Backward number of
bytes in a flow, Forward/Backward number of bytes
used for headers, etc.;
•Packet-based features: Forward/Backward number
of packets in a flow, Forward/Backward length of
packets in a flow (min, max, mean, std), etc.;
•Flow-based features: Length of a flow (mean, max,
etc.);
•Flag-based features: Number of packets with active
TCP/IP flags (FIN, SYN, RST, PUSH, URG, etc.).
6. Experimental Results
Herein we describe our analytical study, which required
the development of a dedicated Python routine, to normal-
ize/balance the datasets and to automate the comparison
among FS algorithms. We have validated the scrutinized
search algorithms using the Correlation-based Feature Se-
lector (CFS) as objective function [151, 152].
The main idea behind CFS is that a good feature sub-
setincludesthosefeaturesthatarehighlycorrelatedwiththe
class,whilebeingstronglyuncorrelatedamongthem. Afor-
mal definition is offered in [153]: a feature Xiis relevant iff
there is some xiandyfor whichp.Xi=xi/>0such that
p.Y=yðXi=xi/‘p.Y=y/: (11)
Namely,Xiis relevant if Yis conditionally dependent on
Xi. Thus, CFS is a filter algorithm that can rank feature
subsets according to a correlation-based heuristic function.
Precisely,givenasubset Sincludingkfeatures,theheuristic
meritMS;kis defined as:
MS;k=krfct
k+k.k* 1/rff; (12)
whererfcis the average value of feature/class correlations,
andrffis the average value of feature/feature correlations.
The numerator of (12) may be seen as an indicator of how
far a set of features is predictive of a class; whereas, the
denominator contains information about how much redun-
dancy there is among features.
Di Mauro et al.: Preprint submitted to Elsevier Page 8 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
(a) Ant (21 fts)
 (b) Scatter (4 fts)
 (c) MO-EA (5 fts)
(d) Ranking (10 fts)
 (e) Cuckoo (7 fts)
 (f) Tabu/LFS (6 fts)
(g) Genetic (27 fts)
 (h) PSO (18 fts)
Figure 2: Correlation maps for different algorithms - DDoS dataset. In parenthesis is
reported the number of features surviving after the FS process.
Our assessment is split into two parts: the first one con-
cernsasingleclass analysis,whereweevaluatedatasetsex-
hibitingdichotomous information (malign/benign); the sec-
ond one is focused on multi class problems, where we eval-
uate the effectiveness of FS in the presence of multiple
classes.
6.1. Single Class Analysis
Let us consider the Distributed Denial of Service
(DDoS) attack which, recently, is also affecting modern
SDN-basednetworks[154,155]. DDoSattacksaredesignedto overwhelm the target network resources by means of a
botnet, namely, a network composed of a large number of
malicious nodes sending tiny packets towards the target, ul-
timately coordinated by a botmaster .
Let us now analyze the results obtained by pre-
processing the DDoS dataset through the set of FS algo-
rithms introduced above. In Fig. 2 we report, for each al-
gorithm, the correlation map corresponding to a graphical
representation of covariance matrices. This representation
embedsthreeimportantpiecesofinformation: i)thenumber
offeaturessurvivingaftertheFSprocessingstep; ii)thetype
Di Mauro et al.: Preprint submitted to Elsevier Page 9 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
0 1 2 3 4 5
Training Size ×104100101102Feat. Sel. Time (sec)MO-EA
Rank
Ant
Tabu
Genetic
Particle Swarm
Cuckoo
Lin.Fwd.Sel.
Scatter
0 1 2 3 4 5
Training Size ×104100101102Training Time (sec)NO Feat. Sel.
MO-EA
Rank
Ant
Tabu
Genetic
Particle Swarm
Cuckoo
Lin.Fwd.Sel.
Scatter
Figure 3: FS times - DDoS dataset (a); Training times - DDoS dataset (b).
offeatures;and iii)therelationshipexistingamongsurviving
features. Thelatteristakenintoaccountbymeansofagray
scale, in which darker shades indicate higher levels of cor-
relation. Thus, each .i;j/“pixel" gives the correlation level
between feature iand featurej. Accordingly, the pixels on
the main diagonal are always black (maximum correlation,
corr= 1), due to the self-correlation. As was to be expected,
highercorrelationarefoundamongthosefeaturesbelonging
to the same family (Time-based, Flow-based, etc.).
Some interesting considerations about the various cor-
relation maps arise. First, the number of features retained
by different algorithms may significantly diverge, which is
due to the specific approaches adopted by each algorithm.
TheGeneticalgorithmistheoneretainingthemostfeatures.
This is to be ascribed to the particular strategy of this al-
gorithm, which strives to escape local optima by applying
themutationoperator,thusallowingtoconsidermorepaths,
namely, more features. Second, some common features re-
tained by all the algorithms can be recognized. For in-
stance,thedestinationportfeatureisalwayspresentsince,in
aDDoSattack, atargetvictimis typicallyreachedonapar-
ticular exposed TCP/UDP port. Moreover, since DDoS at-
tacksarecharacterizedbyalargeamountofsmall-sizepack-
ets, features embodying information about packet lengths
are retained. The difference is that, some algorithms (e.g.
Scatter, MO-EA, Cuckoo, Tabu, LFS) just keep the essen-
tialfeaturesrelatedtopacketlength(e.g. totalpacketlength,
total number of bytes sent in initial window); whereas,
other algorithms (e.g. Ranking, Genetic, PSO, Ant) pre-
fer to retain more features belonging to the same family.
DDoSisalsocharacterizedbysomekindofsynchronization
amongthebots,whicharecoordinatedtolaunchanalmost-
simultaneous attack. This means that time-related features
willoftenprovideusefulinformationtodetectDDoS.Inter-
estingly, the Genetic algorithm retains 5features relating to
the inter-arrival flow times, resulting in a dark gray cluster
at the center of the correlation map (Fig. 2(g)).ItisalsopossibleforDDoSattackstobeevenmoreeffec-
tivethroughthemodificationoftheIPflags(e.g. SYN/RST
flooding). Accordingly, features embodying information
aboutIPflags(e.g. RST-SYN-URGflagcount)areretained
by algorithms such as Ant (Fig. 2(a)), MO-EA (Fig. 2(c)),
Cuckoo (Fig. 2(e)), Genetic (Fig. 2(g)), and PSO (Fig.
2(h)). Let us note that many algorithms opt for selecting
featuresthatareuncorrelatedamongthem(fewdarkgrayor
blackclustersarepresent)sincetheyconveymorevariegated
information.
Let us now analyze some findings obtained from the
time-complexity evaluation. To this aim, we use a PC
equipped with Intel CoreTMi5-7200U CPU@ 2.50GHz
CPUand16GBofRAM.InFig. 3(a),weshowhowtheFS
timevarieswithtrainingsize,fortheDDoSdataset. Nodra-
maticdifferencesareobservedacrossthevariousalgorithms,
even more significantly as the training size grows. Consid-
ering a relatively large training size (with 5104training
instances), FS times range from about 10seconds (Scatter
algorithm) to almost 26seconds (MO-EA algorithm). Sur-
prisingly, the FS times are rather uniform, in spite of the
broadvariationinnumberofretainedfeatures(byeachofthe
algorithms). For instance, remaining in the case of 5104
training instances, Scatter retains the minimum number of
features (4), while Genetic retains the maximum number of
features(27);yetFStimesarecomparable( 16:19and10:18
seconds, respectively). Although it is legitimate to expect
that higher FS time could be justified to produce a more re-
ducedfeaturespace,thescarcecorrelationbetweensuchob-
servables is due to the particular logic implemented in each
FS algorithm.
On the other hand, Fig. 3(b) provides the training times
obtained by applying the J48 benchmark algorithm, down-
stream of the FS processing step. Here, the black line (with
emptycircles)givesthetrainingtimesobtainedwhennoFS
processing is employed. We can observe how FS leads to
significantimprovements,intermsofbothtimesandtrends.
Di Mauro et al.: Preprint submitted to Elsevier Page 10 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
  NO F.S.       MO-EA         Rank          Ant         Tabu        Genetic        PSO        Cuckoo         LFS        Scatter   0.970.9750.980.9850.990.99511.0051.01DDoS Dataset
Accuracy (DDoS)
F-Measure (DDoS)
Accuracy (Benign)
F-Measure (Benign)
(a)
  NO F.S.       MO-EA         Rank          Ant         Tabu        Genetic        PSO        Cuckoo         LFS        Scatter   0.970.9750.980.9850.990.99511.0051.01Portscan Dataset
Accuracy (Portscan)
F-Measure (Portscan)
Accuracy (Benign)
F-Measure (Benign) (b)
  NO F.S.       MO-EA         Rank          Ant         Tabu        Genetic        PSO        Cuckoo         LFS        Scatter   0.970.9750.980.9850.990.99511.005WebAttack Dataset
Accuracy (WebAttack)
F-Measure (WebAttack)
Accuracy (Benign)
F-Measure (Benign)
(c)
  NO F.S.       MO-EA         Rank          Ant         Tabu        Genetic        PSO        Cuckoo         LFS        Scatter   0.970.9750.980.9850.990.99511.005TOR Dataset
Accuracy (TOR)
F-Measure (TOR)
Accuracy (Non TOR)
F-Measure (Non TOR) (d)
Figure 4: Performance in terms of Accuracy/F-Measures for different single class datasets:
DDoS (a), Portscan (b), WebAttack (c), TOR (d).
The black (benchmark) line grows rapidly to almost 80 sec-
onds,whilemostalgorithmspeaktoalmost5seconds,with
theexceptionoftheGeneticalgorithm(yellowline)andthe
Particle Swarm algorithm (light blue line) that take over 10
seconds to complete. This indicates that the FS process, on
the whole, brings gains in the range of about one order of
magnitude,whichmaybecomeevenmoresignificantasthe
dataset grows.
Let us now analyze the performance of the proposed FS
algorithmsintermsofAccuracyandF-Measure. Thesetwo
metrics,widelyusedinthefieldoftrafficclassification[156,
157], are defined as follows:
•Accuracy : the ratio of the correctly predicted obser-
vations to the total observations. This is the most in-
tuitive indicator.
•F-Measure : the weighted average of precision (ratio
ofcorrectlyclassifiedflowsoverallpredictedflowsina class) and recall (ratio of correctly classified flows
overallgroundtruthflowsinaclass). Thisisanindi-
cator of a per-class performance.
To verify that the effectiveness of the FS algorithms is
notlinkedtospecificdatasets,wehaveconsideredthe 4dif-
ferent datasets introduced in Sect. 5(DDoS, Portscan, We-
bAttack, and TOR), reporting our findings in Fig. 4. Just
like for the previous experiments, we have used the tree-
based J 48algorithm as a benchmark. We have adopted a
10-foldcross-validationwhichistypicalinappliedML,and
offersagoodtrade-offbetweentrainingtimeandrobustness.
Noticeably,allFSalgorithmsperformsatisfactorily(bothin
accuracy and F-measure) in comparison to the benchmark
(firstbarsinallthehistograms,labeledas“NOF.S.”)forthe
four datasets.
InsomeinstancestheFSalgorithmsperformedevenbet-
ter than the benchmark (e.g., Rank and Genetic algorithms
in the WebAttack dataset). This can be explained by a phe-
Di Mauro et al.: Preprint submitted to Elsevier Page 11 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
(a) Ant (22 fts)
 (b) Scatter/Tabu (9 fts)
Tot Lenof BwdPktsFwdPktLenStdBwdPktLenMeanInit_Win_bytes_FwdInit_Win_bytes_BwdFlow Pkt/sTot Lenof BwdPktsFwdPktLenStdBwdPktLenMeanInit_Win_bytes_FwdInit_Win_bytes_BwdFlow Pkt/s (c) MO-EA (6 fts)
(d) Ranking (28 fts)
 (e) Cuckoo (17 fts)
Tot Lenof BwdPktsFwdPktLenStdAvgBwdSegmentSizeInit_Win_bytes_FwdInit_Win_bytes_BwdFlow IAT MaxFwdPktLenMaxBwdPktLenStdFlow IAT MinTot Lenof BwdPktsFwdPktLenStd
AvgBwdSegmentSizeInit_Win_bytes_FwdInit_Win_bytes_BwdFlow IAT MaxFwdPktLenMaxBwdPktLenStdFlow IAT Min (f) LFS (9 fts)
(g) Genetic (31 fts)
 (h) PSO (23 fts)
Figure 5: Correlation maps - MultiAndroid dataset. In parenthesis is reported the number
of features surviving after the FS process.
nomenon that is well-known in ML, whereby models based
on too many features may lead to biased classification. On
theotherhand,whenFSmanagestoretainasufficientlyhigh
number of meaningful features, there is a positive effect on
accuracy. This is the case of the Genetic algorithm applied
to the TOR dataset (Fig. 4(d)) that performs better than the
other methods.6.2. Multi Class Analysis
Another fruitful analysis is aimed at evaluating FS al-
gorithms when multi-instance datasets are considered. This
turns out to be particularly useful when it is not possible to
discerndifferenttypesofdatatrafficviasomepre-processing
filter (e.g. IP/Port-based filtering). To assess this case, we
consider two datasets: the MultiAndroid dataset, containing
benigntrafficmixedupwithfivedifferenttypesofAndroid-
based threats; and the DDoS/Portscan dataset, including a
Di Mauro et al.: Preprint submitted to Elsevier Page 12 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
0 1 2 3 4 5
Training Size ×104100101102Feat. Sel. Time (sec)MO-EA
Rank
Ant
Tabu
Genetic
Particle Swarm
Cuckoo
Lin.Fwd.Sel.
Scatter
(a)
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Training Size ×104100101102103Training Time (sec)NO Feat. Sel.
MO-EA
Rank
Ant
Tabu
Genetic
Particle Swarm
Cuckoo
Lin.Fwd.Sel.
Scatter (b)
Figure 6: FS times - MultiAndroid dataset (a); Training times - MultiAndroid dataset (b).
mix of DDoS, Portscan, and benign traffic. The MultiAn-
droiddataset,includesthefollowingtypesofmaligntraffic:
•FakeApp.AL : a scareware hidden inside a fake
Minecraft application, one of the most popular game
applications;
•Android Defender : a malware which, once acci-
dentally downloaded and installed, raises some fake
alerts;
•Gooligam : an insidious malware that has already in-
fected more than 1million Android-based devices,
aimed at stealing Google accounts for Drive, Docs,
Gmail, etc.;
•Feiwo: belonging to the adware family, it acts by
showingadvertisementsinthesystemnotificationbar,
and by sending device GPS coordinates to a remote
server;
•Charger: a ransomware hidden in some Google Play
applications, which gains root privileges and steals
contacts before asking for a ransom.
Let us analyze how FS algorithms impact on the Mul-
tiAndroid dataset in terms of feature correlation referring
first to the panels of Fig. 5. Comparing these results with
the ones of Fig. 2, an interesting difference emerges: all FS
algorithms retain more features w.r.t. the single-class case.
Thisbehavioriscoherentwiththefactthat,todealwithdif-
ferent types of threats (ransomware, adware, malware) we
need more features, to be able to capture this higher vari-
ability. This effect is even more evident in time-based fea-
tures (mainly inter-arrival times) and in size-based features
(mainly packet lengths).
Looking at DDoS, we observe a difference between
single- and multi-class analysis. In the latter, the destina-tion port is not retained as a crucial feature. This is possi-
blybecausemalwaresexploitdifferentmechanismstocreate
damage: rather than directly overwhelming a particular tar-
getport,theyfirstactinthebackground(e.g. bystealingpri-
vacy data) and then produce malicious traffic in egress. On
the other hand, DDoS attacks generate ingress traffic from
the infected device.
It is worth noticing that, when applied to multi-class
problems,allalgorithmshavepreservedtheiroriginallogic.
For instance, with 31 surviving features, the Genetic algo-
rithm is still the algorithm that saves more features, thanks
totheroleplayedbythemutationoperator. Anotherexample
istheMO-EAalgorithmthat,justlikeinthesingle-classex-
periment,retainsthesmallestnumberoffeatures( 6). Thisis
mainly due to the diversity-preservation mechanism, which
forces the selection of a representative subset of the whole
Pareto front. It optimizes conflicting objective functions,
thus few solutions survive.
The time-complexity evaluation is reported in Fig. 6,
which evaluates the usual FS algorithms onto the MultiAn-
droid dataset. FS times exhibit the same order of magni-
tude as in single-class analysis (Figs.3(a)). For a training
size amounting to 5104instances, the fastest algorithm is
Scatter (FS time amounting to 9:541seconds); whereas the
slowest one is MO-EA (FS time amounting to 24:827sec-
onds).
The situation changes dramatically when we consider
training times for the J 48benchmark algorithm (Fig. 6(b)).
Notably, multi-class algorithms are roughly one order of
magnitudeslowerthantheirsingle-classcounterpart. Forin-
stance,letusconsidertheGeneticalgorithm(yellowcurve).
Fora 103trainingsize,GeneticFSreducesthetrainingtime
to1:861seconds, growing to the following (X;Y) points:
(104;10:731); (2 < 104;56:748); (3 < 104;133:346);
(5 < 104;301:997). Thelongertrainingtimesarisefromthe
process of training multiple classes. Nevertheless, signifi-
Di Mauro et al.: Preprint submitted to Elsevier Page 13 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
   NO F.S.       MO-EA        Rank         Ant          Tabu        Genetic       PSO         Cuckoo      LFS/Scat   0.20.30.40.50.60.70.80.9Multi-Class Dataset (Android threats) - Accuracy
FakeAppal
Andr.Defender
Gooligan
Feiwo
Charger
Benign
(a)
   NO F.S.       MO-EA        Rank         Ant          Tabu        Genetic       PSO         Cuckoo      LFS/Scat   0.20.250.30.350.40.450.50.550.6Multi-Class Dataset (Android threats) - F-Measure
FakeAppal
Andr.Defender
Gooligan
Feiwo
Charger
Benign (b)
  NO F.S.       MO-EA         Rank          Ant         Tabu        Genetic        PSO        Cuckoo         LFS        Scatter   0.980.9850.990.99511.005Multi-Class Dataset (DDoS/Portscan) - Accuracy
DDoS
Portscan
Benign
(c)
  NO F.S.       MO-EA         Rank          Ant         Tabu        Genetic        PSO        Cuckoo         LFS        Scatter   0.980.9850.990.99511.005Multi-Class Dataset (DDoS/Portscan) - F-Measure
DDoS
Portscan
Benign (d)
Figure 7: MultiAndroid dataset: Accuracy (a), F-Measure (b); DDoS/Portscan dataset:
Accuracy (c), F-Measure (d);
cant gains are still obtained by all FS algorithms compared
to the “NO F.S.” benchmark, which peaks at 446:329secs.
Turning now to the performance analysis, in Fig. 7
wecomparethetwomulti-classdatasets,MultiAndroidand
DDoS/Portscan,drawingsomeinterestingconsiderations. It
is comparably more difficult to detect Android threats than
DDoS/Portscan attacks - MultiAndroid accuracy is below
0:7and F-Measure is below 0:5. However, this issue is
not generated by the FS processes, since the “NO F.S.” per-
formance is poor too, particularly with the “Benign” class.
This issue arises from two facts. First, mobile network at-
tacks are often accompanied by activities that do not di-
rectly/immediately generate network anomalies. Examples
are ransomware and malware, whereby the anomalies arise
after the user has downloaded the malicious application.
There is typically a lag between infection and anomalies,
as the malicious program initially establishes a secret/silent
communication with a remote server, and then graduallysteals/sends private user data. Another example is adware,
wherethoseannoyingbannersactuallyincurverylittledata,
thusmakingithardtodetectfromtheregulartraffic. Asec-
ond reason for the poor MultiAndroid performance is the
strongsimilarityamongdifferentmalignclasses(e.g.,scare-
ware, adware, ransomware). Similar considerations hold
trueinthecaseinwhichweconsideradatasetincludingWe-
battackandTORtraffic(notreportedforspaceconstraints),
whereby the high similarity between the two classes re-
sulted in poor classification performance. We should how-
ever stress that FS algorithms are still very beneficial, since
thetime-complexitybenefitsidentifiedareachievedwithno
dramatic loss in accuracy.
By contrast, the DDoS/Portscan multi-class case
achieves outstanding performance (Figs.7(c) and (d)). This
is because these types of attacks are radically distinct in
the way they exploit network vulnerabilities: DDoS falls
under the umbrella of volumetric attacks; whereas Portscan
Di Mauro et al.: Preprint submitted to Elsevier Page 14 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
attacks employ monitoring strategies to unveil possible
open ports. In other words, a peculiar symptom of a DDoS
attack is the presence of an exceptionally large number
of connections coming from different nodes and heading
towards one network target’s port. Conversely, a symptom
of Portscan attacks is the presence of just a single node (or
a few nodes in case of simultaneous Portscans) opening a
considerably large number of connections towards multiple
ports of a certain network target. Thus it is relatively easier
to differentiate between these two attacks.
6.3. General Remarks
Overall, we can observe that FS algorithms do lead to
an effective reduction in feature space, ranging from 65~
(Single Class, Genetic) to 95~(Single Class, Scatter) and
from 60~(Multi Class, Genetic) to 92~(Multi Class, MO-
EA). Such feature-space reduction translates into signifi-
cantcomputational-timeimprovements,whichbecomeeven
more remarked as the training size grows. For instance,
with a training set of 50ksamples (single-class DDoS) the
MO-EA algorithm takes 24:8secs to perform FS, while the
trainingtimecomparedtothebenchmarkdropsfrom 72:2to
5:13secs. Atthesametime,performanceisnotsignificantly
degraded by the feature reduction process - accuracy drops
from 0:9993to0:9971. Similar considerations hold for all
other algorithms.
The performed assessment provides invaluable guide-
linesfornetwork/securitymanagementpractitionersdealing
with traffic classification problems. Our evaluation frame-
work aims at weighing the practical benefits of the vari-
ous FS techniques in terms of time-complexity reduction
and performance guarantees. For instance, if we aimed at
minimizing the overall processing time (i.e., FS plus train-
ing times), the Scatter algorithm would be the best choice.
Thisincursatotalprocessingtimeamountingto 14:338sec-
onds for the single-class case (FS= 10:178secs plus train-
ing= 4:16secs),andto 219:963secondsformulti-class(FS=
9:541secsplustraining= 210:422secs). Conversely,theGe-
neticmethodwouldbepreferabletomaximizeperformance.
7. Conclusion and Future Direction
Aprominentresearchdirectionfornetworkintrusionde-
tectionistheadoptionofmachinelearningmethods,partic-
ularly for the detection of anomalous (and often malicious)
network-traffic flows. Looking at the literature, we find am-
ple examples of network classification problems. Yet, little
attention has been turned towards feature selection, which
is an essential classification pre-processing step. We argue
that the main reason for this overlook is that most studies
have been based on the obsolete KDD 99dataset, which in-
cludesfewfeatures,thusmakingFSirrelevant. Ontheother
hand, we consider that modern network engines generate
much richer features (in fact, hundreds of features), which
allowmorefineandgranularnetworktrafficanalyses. How-
ever, this extra capability results into impractical ML train-
ingtimes,makingitnecessarytounderstandhowFSmaybe
realized effectively.To this end, herein we have carried out an experimental
comparativeevaluationofprominentmethods,withtheview
to provide insights as to how the different FS algorithms
perform in the peculiar context of network-traffic classifica-
tion. Our assessment shows how few, relevant features are
retained, but also that the FS reduction process is virtually
lossless, with a significant acceleration of the overall train-
ing process.
To sum up, the novelties of our work are:
i)we carry out an experimental-based review, consider-
ingrecentdatasets(includingDDoS,Portscan,WebAttacks,
and Android threats), as opposed to the obsolete KDD 99
dataset adopted in most literature;
ii)we compare and contrast a representative number
of alternative FS algorithm types, including classic rank-
guided methods (LFS, Ranking), meta-heuristic methods
(Particle Swarm, Tabu, Scatter), nature-inspired methods
(Ant, Cuckoo), and evolutionary methods (Genetic, MO-
EA);
iii)we provide actual experimental results, unveiling
trade-offs between performance (Accuracy/F-Measure) and
computational time, at different scales (training set size).
Ultimately,ouranalysisshowsthebenefitslinkedtoem-
bedding the FS process into network analysis, providing a
valuable tool for identifying the most useful features out of
hundreds of possibilities. This will prove invaluable to the
fieldsofnetworkmanagement,securitymanagement,intru-
sion detection and incident response. We should note that,
the purpose of our comparative evaluation was not to claim
the predominance of some FS algorithms over others but,
rather, to suggest a methodical framework to work with FS.
As a byproduct of our investigation, some interesting
open research directions emerge: i)extending the present
analysisto unsupervised FStechniques,whichwouldbeuse-
ful to deal with datasets lacking class labels, or with new
types of (unknown) malicious traffic - this is the case of so
called zero-day attacks that have no prior information; ii)
consideringthecaseofstreameddataanalysis,whichisnec-
essary when dealing with extremely time-variant streams,
wherebytheFSprocessshouldberepeatedacrosstime(e.g.
byusingamobiletimewindow),soastoperiodicallyupdate
theresultingdatasetwiththefreshestfeatures; iii)designing
routines to automatically manage the best FS strategies to
be applied in accordance to specific criteria (e.g. accuracy
target, latency needs, etc.). Our investigation goes into the
direction of the 6G paradigm that, according to most net-
workscientists,willbecharacterizedbyintelligentresource
management,smartadjustments,andautomaticservicepro-
visioning.
References
[1] F. Camastra and A. Staiano, “Intrinsic dimension estimation: Ad-
vances and open problems,” Information Sciences , vol. 328, pp. 26
– 41, 2016.
[2] F. Camastra, “Data dimensionality estimation methods: a survey,”
Pattern Recognition , vol. 36, no. 12, pp. 2945 – 2954, 2003.
[3] I. Possebon, A. Santos da Silva, L. Zambenedetti Granville,
A. Schaeffer-Filho, and A. Marnerides, “Improved network traffic
Di Mauro et al.: Preprint submitted to Elsevier Page 15 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
classification using ensemble learning,” in 2019 IEEE Symposium
on Computers and Communications (ISCC) , 2019.
[4] F. Grando, L. Zambenedetti Granville, and L. Lamb, “Machine
learninginnetworkcentralitymeasures: Tutorialandoutlook,” ACM
Comput. Surv. , vol. 51, no. 5, pp. 102:1–102:32, 2018.
[5] R. Stadler, R. Pasquini, and V. Fodor, “Learning from network
device statistics,” Journal of Network and Systems Management ,
vol. 25, no. 4, pp. 672–698, 2017.
[6] A. A. Daya, M. A. Salahuddin, N. Limam, and R. Boutaba, “A
graph-based machine learning approach for bot detection,” in 2019
IFIP/IEEESymposiumonIntegratedNetworkandServiceManage-
ment (IM) , 2019, pp. 144–152.
[7] G. Li, X. Hu, X. Shen, X. Chen, and Z. Li, “A novel unsupervised
feature selection method for bioinformatics data sets through fea-
ture clustering,” in 2008 IEEE International Conference on Granu-
lar Computing , 2008, pp. 41–47.
[8] C. Zheng, D. Huang, L. Zhang, and X. Kong, “Tumor clustering
using nonnegative matrix factorization with gene selection,” IEEE
Transactions on Information Technology in Biomedicine , vol. 13,
no. 4, pp. 599–607, 2009.
[9] D. Huang and H. Yu, “Normalized feature vectors: A novel
alignment-free sequence comparison method based on the numbers
of adjacent amino acids,” IEEE/ACM Transactions on Computa-
tionalBiologyandBioinformatics ,vol.10,no.2,pp.457–467,2013.
[10] H.Abusamra,“Acomparativestudyoffeatureselectionandclassifi-
cationmethodsforgeneexpressiondataofglioma,” ProcediaCom-
puter Science , vol. 23, pp. 5 – 14, 2013.
[11] A. Khotanzad and Y. Hong, “Rotation invariant image recognition
using features selected via a systematic method,” Pattern Recogni-
tion, vol. 23, no. 10, pp. 1089 – 1101, 1990.
[12] J. Y. Choi, Y. M. Ro, and K. N. Plataniotis, “Boosting color feature
selection for color face recognition,” IEEE Transactions on Image
Processing , vol. 20, no. 5, pp. 1425–1434, 2011.
[13] A. Goltsev and V. Gritsenko, “Investigation of efficient features for
image recognition by neural networks,” Neural Networks , vol. 28,
pp. 15 – 23, 2012.
[14] D. L. Swets and J. J. Weng, “Using discriminant eigenfeatures for
image retrieval,” IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence , vol. 18, no. 8, pp. 831–836, 1996.
[15] E.Rashedi,H.Nezamabadi-pour,andS.Saryazdi,“Asimultaneous
feature adaptation and feature selection method for content-based
imageretrievalsystems,” Knowledge-BasedSystems ,vol.39,pp.85
– 94, 2013.
[16] Y.Liang,M.Zhang,andW.Browne,“Imagefeatureselectionusing
geneticprogrammingforfigure-groundsegmentation,” Engineering
Applications of Artificial Intelligence , vol. 62, pp. 96–108, 2017.
[17] R. Chatterjee, T. Maitra, S. Hafizul Islam, M. Mehedi Hassan,
A. Alamri, and G. Fortino, “A novel machine learning based fea-
tureselectionformotorimageryeegsignalclassificationininternet
of medical things environment,” Future Generation Computer Sys-
tems, vol. 98, pp. 419–434, 2019.
[18] K. Zhang, Y. Li, P. Scarf, and A. Ball, “Feature selection for high-
dimensional machinery fault diagnosis data using multiple mod-
els and radial basis function networks,” Neurocomputing , vol. 74,
no. 17, pp. 2941 – 2952, 2011.
[19] T.W.Rauber,F.deAssisBoldt,andF.M.Varejao,“Heterogeneous
featuremodelsandfeatureselectionappliedtobearingfaultdiagno-
sis,”IEEETransactionsonIndustrialElectronics ,vol.62,no.1,pp.
637–646, 2015.
[20] D.Lewis,Y.Yang,T.Rose,andF.Li,“Rcv1: Anewbenchmarkcol-
lectionfortextcategorizationresearch,” JournalofMachineLearn-
ing Research , vol. 5, pp. 361–397, 2004.
[21] S. V. Landeghem, T. Abeel, Y. Saeys, and Y. V. de Peer, “Discrim-
inative and informative features for biomolecular text mining with
ensemblefeatureselection,” Bioinformatics ,vol.26,no.18,pp.554–
560, 2010.
[22] M. Labani, P. Moradi, F. Ahmadizar, and M. Jalili, “A novel multi-
variate filter method for feature selection in text classification prob-lems,”Engineering Applications of Artificial Intelligence , vol. 70,
pp. 25–37, 2018.
[23] M.Injadat,A.Moubayed,A.B.Nassif,andA.Shami,“Multi-stage
optimizedmachinelearningframeworkfornetworkintrusiondetec-
tion,”IEEE Transactions on Network and Service Management , pp.
1–1, 2020.
[24] C.Xu,R.Zhang,M.Xie,andL.Yang,“Networkintrusiondetection
systemasaserviceinopenstackcloud,”in 2020InternationalCon-
ference on Computing, Networking and Communications (ICNC) ,
2020, pp. 450–455.
[25] A. Shahraki, M. Abbasi, and O. Haugen, “Boosting algorithms for
network intrusion detection: A comparative evaluation of real ad-
aboost, gentle adaboost and modest adaboost,” Engineering Appli-
cations of Artificial Intelligence , vol. 94, 2020.
[26] M. Di Mauro, M. Longo, F. Postiglione, G. Carullo, and M. Tam-
basco,“ServicefunctionchainingdeployedinanNFVenvironment:
An availability modeling,” in 2017 IEEE Conference on Standards
for Communications and Networking (CSCN) , 2017, pp. 42–47.
[27] M. Di Mauro, M. Longo, F. Postiglione, and M. Tambasco, “Avail-
ability modeling and evaluation of a network service deployed via
NFV,”in DigitalCommunication.TowardsaSmartandSecureFu-
ture Internet , 2017, pp. 31–44.
[28] M.DiMauro,M.Longo,andF.Postiglione,“Availabilityevaluation
of multi-tenant service function chaining infrastructures by multi-
dimensional universal generating function,” IEEE Transactions on
Services Computing , pp. 1–1, 2018.
[29] M. Di Mauro, M. Longo, F. Postiglione, R. Restaino, and M. Tam-
basco,“Availabilityevaluationofthevirtualizedinfrastructureman-
ager in network function virtualization environments,” in Proc. of
the 26th European Safety and Reliability Conference, ESREL 2016 ,
2017, pp. 2591–2596.
[30] M.DiMauro,M.Longo,andF.Postiglione,“Reliabilityanalysisof
thecontrollerarchitectureinsoftwaredefinednetworks,”in Proc.of
the 26th European Safety and Reliability Conference, ESREL 2015 ,
2015, pp. 1503–1510.
[31] M. Di Mauro, G. Galatro, M. Longo, F. Postiglione, and M. Tam-
basco, “Availability evaluation of a virtualized IP Multimedia Sub-
systemfor5Gnetworkarchitectures,”in Proc.ofthe26thEuropean
Safety and Reliability Conference, ESREL 2017 , 2017, pp. 2203–
2210.
[32] ——,“ComparativeperformabilityassessmentofSFCs: Thecaseof
containerizedIPMultimediaSubsystem,” IEEETrans.Netw.Service
Manag., 2020.
[33] ——, “Performability management of softwarized IP Multimedia
Subsystem,” in IEEE/IFIP Network Operations and Management
Symposium, 2020 , 2020, pp. 1–6.
[34] M. Di Mauro, G. Galatro, M. Longo, A. Palma, F. Postiglione,
and M. Tambasco, “Automated generation of availability models
for SFCs: The case of virtualized IP Multimedia Subsystem,” in
IEEE/IFIPNetworkOperationsandManagementSymposium,2020 ,
2020, pp. 1–6.
[35] V.Matta,M.DiMauro,andM.Longo,“Botnetidentificationinran-
domizedDDoSattacks,”in Proceedingsofthe24thEuropeanSignal
Processing Conference , 2016, pp. 2260–2264.
[36] ——, “Botnet identification in multi-clustered DDoS attacks,” in
2017 25th European Signal Processing Conference (EUSIPCO) ,
2017, pp. 2171–2175.
[37] P. Addesso, M. Cirillo, M. Di Mauro, and V. Matta, “ADVoIP: Ad-
versarialdetectionofencryptedandconcealedvoip,” IEEETransac-
tions on Information Forensics and Security , vol. 15, pp. 943–958,
2020.
[38] V.Matta,M.DiMauro,M.Longo,andA.Farina,“Cyber-threatmit-
igationexploitingthebirth–death–immigrationmodel,” IEEETrans-
actions on Information Forensics and Security , vol. 13, no. 12, pp.
3137–3152, 2018.
[39] W. Cerroni, G. Moro, R. Pasolini, and M. Ramilli, “Decentralized
detection of network attacks through p2p data clustering of snmp
data,”Computers & Security , vol. 52, pp. 1 – 16, 2015.
Di Mauro et al.: Preprint submitted to Elsevier Page 16 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
[40] ——,“Networkattackdetectionbasedonpeer-to-peerclusteringof
snmpdata,”in LectureNotesoftheInstituteforComputerSciences ,
vol. 22, 2009.
[41] W. Cerroni, G. Moro, T. Pirini, and M. Ramilli, “Peer-to-peer data
miningclassifiersfordecentralizeddetectionofnetworkattacks,”in
Proceedings of the Twenty-Fourth Australasian Database Confer-
ence - Volume 137 , 2013, pp. 101–107.
[42] “The CSE-CIC-IDS2018 Dataset,” https://github.com/alekzandr/
flowmeter, accessed: 2020-10-01.
[43] G.ChandrashekarandF.Sahin,“Asurveyonfeatureselectionmeth-
ods,”Computers & Electrical Engineering , vol. 40, no. 1, pp. 16 –
28, 2014.
[44] S. Khalid, T. Khalil, and S. Nasreen, “A survey of feature selection
andfeatureextractiontechniquesinmachinelearning,”in 2014Sci-
ence and Information Conference , 2014, pp. 372–378.
[45] L. C. Molina, L. Belanche, and A. Nebot, “Feature selection algo-
rithms: a survey and experimental evaluation,” in 2002 IEEE Inter-
nationalConferenceonDataMining,2002.Proceedings. ,2002,pp.
306–313.
[46] L. Yu and H. Liu, “Efficient feature selection via analysis of rele-
vanceandredundancy,” TheJournalofMachineLearningResearch ,
vol. 5, pp. 1205–1224, 2004.
[47] F. Camastra and A. Vinciarelli, Feature Extraction Methods and
Manifold Learning Methods . Springer London, 2008.
[48] A. L. Blum and P. Langley, “Selection of relevant features and ex-
amples in machine learning,” Artificial Intelligence , vol. 97, no. 1,
pp. 245 – 271, 1997.
[49] M.DashandH.Liu,“Featureselectionforclassification,” Intelligent
Data Analysis , vol. 1, no. 1, pp. 131 – 156, 1997.
[50] S.Alelyani,J.Tang,andH.Liu,“Featureselectionforclustering: A
review,” in Data Clustering: Algorithms and Applications , 2013.
[51] L. Rendell and R. Seshu, “Learning hard concepts through con-
structive induction: Framework and rationale,” in Proceedings of a
Workshop on Computational Learning Theory and Natural Learn-
ing Systems (Vol. 1) : Constraints and Prospects: Constraints and
Prospects , 1994, pp. 83–141.
[52] L.Talavera,“Anevaluationoffilterandwrappermethodsforfeature
selection in categorical clustering,” in Advances in Intelligent Data
Analysis VI , 2005, pp. 440–451.
[53] M. A. Hall and L. A. Smith, “Feature selection for machine learn-
ing: Comparing a correlation-based filter approach to the wrapper,”
inProceedings of the Twelfth International Florida Artificial Intel-
ligence Research Society Conference , 1999, pp. 235–239.
[54] R. Kohavi and G. H. John, “Wrappers for feature subset selection,”
Artificial Intelligence , vol. 97, no. 1, pp. 273 – 324, 1997.
[55] I. Guyon, “An introduction to variable and feature selection,” Jour-
nal of Machine Learning Research , vol. 3, pp. 1157–1182, 2003.
[56] J. C. Ang, A. Mirzal, H. Haron, and H. N. A. Hamed, “Supervised,
unsupervised, and semi-supervised feature selection: A review on
geneselection,” IEEE/ACMTransactionsonComputationalBiology
and Bioinformatics , vol. 13, no. 5, pp. 971–989, 2016.
[57] J.Dromard,G.Roudière,andP.Owezarski,“Onlineandscalableun-
supervised network anomaly detection method,” IEEE Trans. Netw.
Service Manag. , vol. 14, no. 1, pp. 34–47, 2017.
[58] W. Wang, Y. He, J. Liu, and S. Gombault, “Constructing important
features from massive network traffic for lightweight intrusion de-
tection,”IETInformationSecurity ,vol.9,no.6,pp.374–379,2015.
[59] T. Janarthanan and S. Zargari, “Feature selection in unsw-nb15 and
kddcup’99 datasets,” in 2017 IEEE 26th International Symposium
on Industrial Electronics (ISIE) , 2017, pp. 1881–1886.
[60] K.El-Khatib,“Impactoffeaturereductionontheefficiencyofwire-
lessintrusiondetectionsystems,” IEEETransactionsonParalleland
Distributed Systems , vol. 21, no. 8, pp. 1143–1149, 2010.
[61] Y. Chen, Y. Li, X. Cheng, and L. Guo, “Survey and taxonomy of
feature selection algorithms in intrusion detection system,” in Pro-
ceedingsoftheSecondSKLOISConferenceonInformationSecurity
and Cryptology , 2006, pp. 153–167.
[62] A.Nisioti,A.Mylonas,P.D.Yoo,andV.Katos,“Fromintrusionde-tection to attacker attribution: A comprehensive survey of unsuper-
vised methods,” IEEE Communications Surveys Tutorials , vol. 20,
no. 4, pp. 3369–3388, 2018.
[63] F. Iglesias and T. Zseby, “Analysis of network traffic features for
anomaly detection,” Machine Learning , vol. 101, no. 1, pp. 59–84,
2015.
[64] R.Singh,H.Kumar,andR.K.Singla,“Analysisoffeatureselection
techniques for network traffic dataset,” in 2013 International Con-
ference on Machine Intelligence and Research Advancement , 2013,
pp. 42–46.
[65] M. Bahrololum, E. Salahi, and M. Khaleghi, “Machine learning
techniques for feature reduction in intrusion detection systems: A
comparison,” in Progress in Computing, Analytics and Networking.
Advances in Intelligent Systems and Computing , 2009, pp. 1091–
1095.
[66] Y.Dhote,S.Agrawal,andA.J.Deen,“Asurveyonfeatureselection
techniques for internet traffic classification,” in 2015 International
ConferenceonComputationalIntelligenceandCommunicationNet-
works (CICN) , 2015, pp. 1375–1380.
[67] A. L. Buczak and E. Guven, “A survey of data mining and machine
learningmethodsforcybersecurityintrusiondetection,” IEEECom-
municationsSurveysTutorials ,vol.18,no.2,pp.1153–1176,2016.
[68] P. Mishra, V. Varadharajan, U. Tupakula, and E. S. Pilli, “A de-
tailed investigation and analysis of using machine learning tech-
niques for intrusion detection,” IEEE Communications Surveys Tu-
torials, vol. 21, no. 1, pp. 686–728, 2019.
[69] M. Di Mauro and C. Di Sarno, “Improving SIEM capabilities
through an enhanced probe for encrypted skype traffic detection,”
Journal of Information Security and Applications , vol. 38, pp. 85–
95, 2018.
[70] M. Di Mauro and M. Longo, “Skype traffic detection: A decision
theory based tool,” in 2014 International Carnahan Conference on
Security Technology (ICCST) , 2014, pp. 1–6.
[71] M. Di Mauro and C. Di Sarno, “A framework for internet data real-
time processing: A machine-learning approach,” in 2014 Interna-
tionalCarnahanConferenceonSecurityTechnology(ICCST) ,2014,
pp. 1–6.
[72] F. Cauteruccio, G. Fortino, A. Guerrieri, A. Liotta, D. Mocanu,
C. Perra, G. Terracina, and M. Torres Vega, “Short-long term
anomaly detection in wireless sensor networks based on machine
learning and multi-parameterized edit distance,” Information Fu-
sion, vol. 52, pp. 13 – 30, 2019.
[73] M. Di Mauro, G. Galatro, and A. Liotta, “Experimental review of
neural-based approaches for network intrusion management,” IEEE
Transactions on Network and Service Management , pp. 1–1, 2020.
[74] M. Di Mauro and M. Longo, “Revealing encrypted webrtc traffic
viamachinelearningtools,”in 201512thInternationalJointConfer-
enceone-BusinessandTelecommunications(ICETE) ,vol.04,2015,
pp. 259–266.
[75] H. Benaddi, K. Ibrahimi, and A. Benslimane, “Improving the in-
trusion detection system for nsl-kdd dataset based on pca-fuzzy
clustering-knn,” in 2018 6th International Conference on Wireless
Networks and Mobile Communications (WINCOM) , 2018, pp. 1–6.
[76] W. Wang, X. Du, and N. Wang, “Building a cloud ids using an ef-
ficient feature selection method and svm,” IEEE Access , vol. 7, pp.
1345–1354, 2019.
[77] S.M.KasongoandY.Sun,“Adeeplearningmethodwithfilterbased
feature engineering for wireless intrusion detection system,” IEEE
Access, vol. 7, pp. 38597–38607, 2019.
[78] K. Wu, Z. Chen, and W. Li, “A novel intrusion detection model for
a massive network using convolutional neural networks,” IEEE Ac-
cess, vol. 6, pp. 50850–50859, 2018.
[79] M. A. Ambusaidi, X. He, P. Nanda, and Z. Tan, “Building an in-
trusion detection system using a filter-based feature selection algo-
rithm,”IEEETransactionsonComputers ,vol.65,no.10,pp.2986–
2998, 2016.
[80] K. A. Taher, B. Mohammed Yasin Jisan, and M. M. Rahman, “Net-
work intrusion detection using supervised machine learning tech-
Di Mauro et al.: Preprint submitted to Elsevier Page 17 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
nique with feature selection,” in 2019 International Conference on
Robotics,Electrical and Signal Processing Techniques (ICREST) ,
2019, pp. 643–646.
[81] J. Woo, J. Song, and Y. Choi, “Performance enhancement of deep
neural network using feature selection and preprocessing for intru-
sion detection,” in 2019 International Conference on Artificial In-
telligence in Information and Communication (ICAIIC) , 2019, pp.
415–417.
[82] F. Amiri, M. R. Yousefi, C. Lucas, A. Shakery, and N. Yazdani,
“Mutual information-based feature selection for intrusion detection
systems,” Journal of Network and Computer Applications , vol. 34,
no. 4, pp. 1184 – 1199, 2011.
[83] M. Tavallaee, E. Bagheri, W. Lu, and A. A. Ghorbani, “A detailed
analysis of the kdd cup 99 data set,” in 2009 IEEE Symposium on
Computational Intelligence for Security and Defense Applications ,
2009, pp. 1–6.
[84] N.MoustafaandJ.Slay,“Unsw-nb15: acomprehensivedatasetfor
networkintrusiondetectionsystems(unsw-nb15networkdataset),”
in2015MilitaryCommunicationsandInformationSystemsConfer-
ence (MilCIS) , 2015, pp. 1–6.
[85] N.Moustafa,J.Slay,andG.Creech,“Novelgeometricareaanalysis
techniqueforanomalydetectionusingtrapezoidalareaestimationon
large-scalenetworks,” IEEETransactionsonBigData ,vol.5,no.4,
pp. 481–494, 2019.
[86] Doreswamy, M. K. Hooshmand, and I. Gad, “Feature selection ap-
proach using ensemble learning for network anomaly detection,”
CAAI Transactions on Intelligence Technology , vol. 5, no. 4, pp.
283–293, 2020.
[87] A. Binbusayyis and T. Vaiyapuri, “Identifying and benchmarking
key features for cyber intrusion detection: An ensemble approach,”
IEEE Access , vol. 7, pp. 106495–106513, 2019.
[88] “canadian institute for cybersecurity,” https://www.unb.ca/cic/, ac-
cessed: 2020-10-01.
[89] M.A.HallandG.Holmes,“Benchmarkingattributeselectiontech-
niquesfordiscreteclassdatamining,” IEEETransactionsonKnowl-
edge and Data Engineering , vol. 15, no. 6, pp. 1437–1447, 2003.
[90] R. B and G. S, “An intelligent fuzzy rule based feature selection for
effective intrusion detection,” in 2018 International Conference on
RecentTrendsinAdvanceComputing(ICRTAC) ,2018,pp.206–211.
[91] V. S. Takkellapati and G. V. Prasad, “Network intrusion detection
system based on feature selection and triangle area support vector
machine,” International Journal of Engineering Trends and Tech-
nology, vol. 3, no. 4, pp. 466–470, 2012.
[92] S.Ganapathy,K.Kulothungan,P.Yogesh,andA.Kannan,“Anintel-
ligentintrusiondetectionsystemforadhocnetworks,”in IETChen-
nai3rdInternationalonSustainableEnergyandIntelligentSystems
(SEISCON 2012) , 2012, pp. 1–5.
[93] J.M.HernándezJimà ©nezandK.Goseva-Popstojanova,“Theef-
fectonnetworkflows-basedfeaturesandtrainingsetsizeonmalware
detection,”in 2018IEEE17thInternationalSymposiumonNetwork
Computing and Applications (NCA) , 2018, pp. 1–9.
[94] P. Singh and A. Tiwari, “An efficient approach for intrusion detec-
tion in reduced features of kdd99 using id3 and classification with
knnga,” in 2015 Second International Conference on Advances in
Computing and Communication Engineering , 2015, pp. 445–452.
[95] M. Gutlein, E. Frank, M. Hall, and A. Karwath, “Large-scale at-
tributeselectionusingwrappers,”in 2009IEEESymposiumonCom-
putational Intelligence and Data Mining , 2009, pp. 332–339.
[96] W.Wang,X.Wang,D.Feng,J.Liu,Z.Han,andX.Zhang,“Explor-
ing permission-induced risk in android applications for malicious
applicationdetection,” IEEETransactionsonInformationForensics
and Security , vol. 9, no. 11, pp. 1869–1882, 2014.
[97] I. Finizio, C. Mazzariello, and C. Sansone, “Combining genetic-
basedmisuseandanomalydetectionforreliablydetectingintrusions
in computer networks,” in Proceedings of the 13th International
Conference on Image Analysis and Processing , 2005, pp. 66–74.
[98] Y. Liu, Z. Xu, J. Yang, L. Wang, C. Song, and K. Chen, “A novel
meta-heuristic-based sequential forward feature selection approachfor anomaly detection systems,” in 2016 International Conference
onNetworkandInformationSystemsforComputers(ICNISC) ,2016,
pp. 218–227.
[99] F.Glover,“Futurepathsforintegerprogrammingandlinkstoartifi-
cialintelligence,” Computers&OperationsResearch ,vol.13,no.5,
pp. 533 – 549, 1986.
[100] G.F.W.andL.M., TabuSearch . NewYork,USA:Springer,1997.
[101] C. Rego and B. Alidaee, Metaheuristic Optimization via Memory
and Evolution: Tabu Search and Scatter Search (Operations Re-
search/Computer Science Interfaces Series) . Berlin, Heidelberg:
Springer-Verlag, 2005.
[102] A.-R.Hedar,J.Wang,andM.Fukushima,“Tabusearchforattribute
reduction in rough set theory,” Soft Computing , vol. 12, no. 9, pp.
909–918, 2008.
[103] H.Mohamadi,J.Habibi,andH.Saadi,“Intrusiondetectionincom-
puter networks using tabu search based fuzzy system,” in 2008 7th
IEEE International Conference on Cybernetic Intelligent Systems ,
2008, pp. 1–6.
[104] W.Jian-guang,T.Ran,andL.Zhi-Yong,“Animprovingtabusearch
algorithmforintrusiondetection,”in 2011ThirdInternationalCon-
ference on Measuring Technology and Mechatronics Automation ,
vol. 1, 2011, pp. 435–439.
[105] Y. Chen, L. Dai, and X. Cheng, “Gats-c4.5: An algorithm for opti-
mizing features in flow classification,” in 2008 5th IEEE Consumer
Communications and Networking Conference , 2008, pp. 466–470.
[106] K.Bakour,G.S.Das,andH.M.Unver,“Anintrusiondetectionsys-
tembasedonahybridtabu-geneticalgorithm,”in 2017International
Conference on Computer Science and Engineering (UBMK) , 2017,
pp. 215–220.
[107] XiaocongZ.,DonglingL.,andYangY.,“Improvedincrementalsup-
portvectormachinewithhybridfeatureselectionfornetworkintru-
sion detection,” in 2013 International Conference on Information
and Network Security (ICINS 2013) , 2013, pp. 1–6.
[108] F.Glover,“Heuristicsforintegerprogrammingusingsurrogatecon-
straints,” Decision Sciences , vol. 8, pp. 156–166, 1977.
[109] Z. Ugray, L. Lasdon, J. Plummer, F. Glover, J. Kelly, and R. Marti,
“Scatter search and local nlp solvers: a multistart framework for
globaloptimization,” InformsJournalonComputing ,vol.19,no.3,
pp. 328 – 340, 2007.
[110] J. Wang, A. R. Hedar, S. Wang, and J. Ma, “Rough set and scat-
ter search metaheuristic based feature selection for credit scoring,”
Expert Systems with Applications , vol. 39, no. 6, pp. 6123 – 6128,
2012.
[111] F. G. Lopez, M. G. Torres, B. M. Batista, J. A. M. Perez, and J. M.
Moreno-Vega,“Solvingfeaturesubsetselectionproblembyaparal-
lel scatter search,” European Journal of Operational Research , vol.
169, no. 2, pp. 477 – 489, 2006.
[112] E. Duman and M. H. Ozcelik, “Detecting credit card fraud by ge-
netic algorithm and scatter search,” Expert Systems with Applica-
tions, vol. 38, no. 10, pp. 13057 – 13063, 2011.
[113] D.ByersandN.Shahmehri,“Prioritisationandselectionofsoftware
security activities,” in 2009 International Conference on Availabil-
ity, Reliability and Security , 2009, pp. 201–207.
[114] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Pro-
ceedings of ICNN’95 - International Conference on Neural Net-
works, vol. 4, 1995, pp. 1942–1948 vol.4.
[115] Q.Yao,J.Cai,andJ.Zhang,“Simultaneousfeatureselectionandls-
svmparametersoptimizationalgorithmbasedonPSO,”in 2009WRI
WorldCongressonComputerScienceandInformationEngineering ,
vol. 5, 2009, pp. 723–727.
[116] W.Hu,J.Gao,Y.Wang,O.Wu,andS.Maybank,“Onlineadaboost-
basedparameterizedmethodsfordynamicdistributednetworkintru-
sion detection,” IEEE Transactions on Cybernetics , vol. 44, no. 1,
pp. 66–82, 2014.
[117] H.Li,W.Guo,G.Wu,andY.Li,“ARF-PSObasedhybridfeaturese-
lectionmodelinintrusiondetectionsystem,”in 2018IEEEThirdIn-
ternationalConferenceonDataScienceinCyberspace(DSC) ,2018,
pp. 795–802.
Di Mauro et al.: Preprint submitted to Elsevier Page 18 of 19Supervised Feature Selection Techniques in Network Intrusion Detection: a Critical Review
[118] S. Fong, R. Wong, and A. V. Vasilakos, “Accelerated PSO swarm
search feature selection for data stream mining big data,” IEEE
TransactionsonServicesComputing ,vol.9,no.1,pp.33–45,2016.
[119] M.Dorigo,V.Maniezzo,andA.Colorni,“Antsystem: optimization
by a colony of cooperating agents,” IEEE Transactions on Systems,
Man, and Cybernetics, Part B (Cybernetics) , vol. 26, no. 1, pp. 29–
41, 1996.
[120] T. Mehmood and H. B. M. Rais, “Svm for network anomaly detec-
tion using aco feature subset,” in 2015 International Symposium on
MathematicalSciencesandComputingResearch(iSMSC) ,2015,pp.
121–126.
[121] S.Hardeand V. Sahare,“Designandimplementationof acofeature
selection algorithm for data stream mining,” in 2016 International
ConferenceonAutomaticControlandDynamicOptimizationTech-
niques (ICACDOT) , 2016, pp. 1047–1051.
[122] H. Peng, C. Ying, S. Tan, B. Hu, and Z. Sun, “An improved feature
selectionalgorithmbasedonantcolonyoptimization,” IEEEAccess ,
vol. 6, pp. 69203–69209, 2018.
[123] X. Yang, Cuckoo Search and Firefly Algorithm: Theory and Appli-
cations. London, UK: Springer, 2013.
[124] Z.Li,Y.Su,andQ.Han,“Intrusiondetectionbasedonpcaandfuzzy
clustering optimized by cs,” in 2017 Chinese Automation Congress
(CAC), 2017, pp. 6334–6339.
[125] E. J. L. D. Asir Antony Gnana Singh, R. Priyadharshini, “Cuckoo
optimisationbasedintrusiondetectionsystemforcloudcomputing,”
InternationJournalofComputerNetworkandInformationSecurity ,
vol. 11, pp. 42–49, 2018.
[126] K. Rithesh, “Anomaly-based nids using artificial neural networks
optimised with cuckoo search optimizer,” in Emerging Research in
Electronics, Computer Science and Technology , 2019, pp. 23–35.
[127] W. Niu, X. Zhang, G. Yang, Z. Ma, and Z. Zhuo, “Phishing emails
detection using cs-svm,” in 2017 IEEE International Symposium
on Parallel and Distributed Processing with Applications and 2017
IEEEInternationalConferenceonUbiquitousComputingandCom-
munications (ISPA/IUCC) , 2017, pp. 1054–1059.
[128] M.Redmond,S.Salesi,andG.Cosma,“Anovelapproachbasedon
anextendedcuckoosearchalgorithmfortheclassificationoftweets
which contain emoticon and emoji,” in 2017 2nd International
Conference on Knowledge Engineering and Applications (ICKEA) ,
2017, pp. 13–19.
[129] I.H.Abdulqadder,D.Zou,I.T.Aziz,B.Yuan,andW.Li,“Secsdn-
cloud: Defeatingvulnerableattacksthroughsecuresoftware-defined
networks,” IEEE Access , vol. 6, pp. 8292–8301, 2018.
[130] T. Weise, “Global optimization algorithms – theory and applica-
tion,” 2009.
[131] W.W.BledsoeandI.Browning,“Patternrecognitionandreadingby
machine,” in Papers Presented at the December 1-3, 1959, Eastern
Joint IRE-AIEE-ACM Computer Conference , 1959, pp. 225–232.
[132] H.J.Bremermann,“Optimizationthroughevolutionandrecombina-
tion,” inSelf-Organizing Systems , M. C. Yovits, G. T. Jacobi, , and
G. D. Goldstein, Eds. Spartan Books, 1962.
[133] J.Holland,“Outlineforalogicaltheoryofadaptivesystems,” Jour-
nal of ACM , vol. 9, no. 3, pp. 297–314, 1962.
[134] ——, AdaptationinNaturalandArtificialSystems: AnIntroductory
AnalysiswithApplicationstoBiology,ControlandArtificialIntelli-
gence. Cambridge, MA, USA: MIT Press, 1992.
[135] C.Coello,“Anupdatedsurveyofga-basedmultiobjectiveoptimiza-
tion techniques,” ACM Computing Surveys , vol. 32, no. 2, pp. 109–
143, 2000.
[136] D. Goldberg, Genetic Algorithms in Search, Optimization and Ma-
chineLearning ,1sted. Boston,MA,USA:Addison-WesleyLong-
man Publishing Co., Inc., 1989.
[137] S. Guha, S. S. Yau, and A. B. Buduru, “Attack detection in cloud
infrastructures using artificial neural network with genetic feature
selection,” in 2016 IEEE 14th Intl Conf on Dependable, Autonomic
and Secure Computing , 2016, pp. 414–419.
[138] B. Senthilnayaki, K. Venkatalakshmi, and A. Kannan, “An intelli-
gentintrusiondetectionsystemusinggeneticbasedfeatureselectionandmodifiedJ48decisiontreeclassifier,”in 2013FifthInternational
Conference on Advanced Computing (ICoAC) , 2013, pp. 1–7.
[139] ——, “Intrusion detection using optimal genetic feature selection
and svm based classifier,” in 2015 3rd International Conference on
SignalProcessing,CommunicationandNetworking(ICSCN) ,2015,
pp. 1–4.
[140] H.GharaeeandH.Hosseinvand,“Anewfeatureselectionidsbased
ongeneticalgorithmandsvm,”in 20168thInternationalSymposium
on Telecommunications (IST) , 2016, pp. 139–144.
[141] P. Tao, Z. Sun, and Z. Sun, “An improved intrusion detection al-
gorithm based on GA and SVM,” IEEE Access , vol. 6, pp. 13624–
13631, 2018.
[142] C.M.FonsecaandP.J.Fleming,“Anoverviewofevolutionaryalgo-
rithms in multiobjective optimization,” Evolutionary Computation ,
vol. 3, no. 1, pp. 1–16, 1995.
[143] F. Jimenez, G. Sanchez, J. Garcia, G. Sciavicco, and L. Miralles,
“Multi-objectiveevolutionaryfeatureselectionforonlinesalesfore-
casting,”Neurocomputing , vol. 234, pp. 75 – 92, 2017.
[144] M.S.AliakbarianandA.Fanian,“Internettrafficclassificationusing
moeaandonlinerefinementinvotingonensemblemethods,”in 2013
21st Iranian Conference on Electrical Engineering (ICEE) , 2013,
pp. 1–6.
[145] Y.Zhu,J.Liang,J.Chen,andZ.Ming,“Animprovednsga-iiialgo-
rithm for feature selection used in intrusion detection,” Knowledge-
Based Systems , vol. 116, pp. 74–85, 2017.
[146] “Buildinganeffectiveintrusiondetectionsystemusingunsupervised
feature selection in multi-objective optimization framework,” arxiv.
org/pdf/1905.06562.pdf, accessed: 2020-10-01.
[147] P. Ducange, G. Mannara, F. Marcelloni, R. Pecori, and M. Vecchio,
“A novel approach for internet traffic classification based on multi-
objectiveevolutionaryfuzzyclassifiers,”in 2017IEEEInternational
Conference on Fuzzy Systems (FUZZ-IEEE) , 2017, pp. 1–6.
[148] “The KDD99 Dataset,” http://kdd.ics.uci.edu/databases/kddcup99\
/kddcup99.html, accessed: 2020-10-01.
[149] G.Draper-Gil,A.H.Lashkari,M.S.I.Mamun,andA.A.Ghorbani,
“Characterization of encrypted and vpn traffic using time-related
features,” in International Conference on Information Systems Se-
curity and Privacy , 2016.
[150] “Cicflowmeter network analyzer,” https://www.unb.ca/cic/datasets/
ids-2018.html, accessed: 2020-10-01.
[151] “Correlation-based feature selection for machine learning,” cs.
waikato.ac.nz/~mhall/thesis.pdf, accessed: 2020-10-01.
[152] F. Fabris, A. A. Freitas, and J. M. A. Tullet, “An extensive empiri-
calcomparisonofprobabilistichierarchicalclassifiersindatasetsof
ageing-related genes,” IEEE/ACM Transactions on Computational
Biology and Bioinformatics , vol. 13, no. 6, pp. 1045–1058, 2016.
[153] R. Kohavi and G. H. John, “Wrappers for feature subset selection,”
Artificial Intelligence , vol. 97, no. 1, pp. 273 – 324, 1997.
[154] L. Zhu, X. Tang, M. Shen, X. Du, and M. Guizani, “Privacy-
preserving ddos attack detection using cross-domain traffic in soft-
ware defined networks,” IEEE Journal on Selected Areas in Com-
munications , vol. 36, no. 3, pp. 628–643, 2018.
[155] K. Kalkan, L. Altay, G. Gür, and F. Alagöz, “Jess: Joint entropy-
basedddosdefenseschemeinsdn,” IEEEJournalonSelectedAreas
in Communications , vol. 36, no. 10, pp. 2358–2372, 2018.
[156] H. Kim, K. Claffy, M. Fomenkov, D. Barman, M. Faloutsos, and
K. Lee, “Internet traffic classification demystified: Myths, caveats,
and the best practices,” in Proceedings of the 2008 ACM CoNEXT
Conference , 2008, pp. 11:1–11:12.
[157] J. Zhang, Y. Xiang, Y. Wang, W. Zhou, Y. Xiang, and Y. Guan,
“Network traffic classification using correlation information,” IEEE
TransactionsonParallelandDistributedSystems ,vol.24,no.1,pp.
104–117, 2013.
Di Mauro et al.: Preprint submitted to Elsevier Page 19 of 19