Malware Detection based on HTTPS Characteristic via Machine
Learning
Paul Calderon
1
, Hirokazu Hasegawa
2
, Yukiko Yamaguchi
2
and Hajime Shimada
2
1
Ensimag, Grenoble INP, France
2
Information Technology Center, Nagoya University, Japan
Keywords:
Security, Malware, Infection Detection, Machine Learning.
Abstract:
One of the major threat in today world are malwares that can infect computers. In order to prevent infection
antimalwares softwares are installed but if the malware it is not detected at the installation it will probably
never be detected. Behavioural analysis is necessary. Most of nowadays malwares connect to C&C servers
by utilizing HTTP or HTTPS in order to receive orders. In this paper a method of behavioural analysis focus
on the observation on HTTP and HTTPS network packets will be presented. This analysis is made by using
machine learning. We evaluated our method by using 10-fold cross validations. The experimental result shows
that precisions and recalls are more than 96% in average.
1 INTRODUCTION
Executing malicious programs on a computer is one
of the major threat and attack observed against infor-
matics systems. In order to respond to this attacks, it
become common to users to install antimalware soft-
wares. Different methods are used by this softwa-
res to detect malicious programs, for example static
analysis based on the signature of files or dynamic
analysis using ”sandbox” that try to detect suspici-
ous behaviour. One way usually used to avoid static
analysis detection is the encryption of the program.
However this is usually not useful against dynamic
analysis. Techniques exist to trick behavioural ana-
lysis, some of this techniques are relatively easy to
use. All of this detection methods are made to pre-
vent machine infection. Indeed early detection is very
important, if a malware is detected after the infection,
it could have already achieve is goal such as stealing
data, mining bitcoins, etc. However it is not always
possible to detect malwares in advance, some of them
can avoid detection techniques. The malware could
have been installed because of human error, or be-
cause the malicious behaviour is delayed. For this
reasons it is also important to detect malwares once
the machines are infected. This can be done by ana-
lysing permanently the behaviour of program used on
computer. But it uses resources from the computer.
Most of virus are now communicating on the Internet
with the attacker through a C&C server. Analysing
traffic network can be an option for discovering sus-
picious behaviour and detect infected machines. For
example using a firewall blocking some protocols can
be a possibility but most of malwares communicate
through HTTP or HTTPS that are daily used while
surfing on the web. It is then important to detect
patterns on HTTP and HTTPS packets that allow to
discriminate infected machines network packets from
legitimate packets. Finding the good pattern can be
tough, we are going to use machine learning to au-
tomatize the network analysis. One article presented
at the ICOIN 2017(Ogawa et al., ) shows that ana-
lysing HTTP traffic is powerful method on detecting
malware infected machine. As being able to detect
malware infected machine with a precision and recall
superior of 96%.
2 MALWARE’S
COMMUNICATION
The first malwares were only doing harm to infected
computers, without the ability to change their beha-
viour depending on what the attacker wanted. But
having the possibility to control the malware at dis-
tance increase the power of threat of malwares, with
the possibility to give them orders that modify their
behaviour depending on the situation. In order to give
order to malwares hackers use what it is called com-
410
Calderon, P., Hasegawa, H., Yamaguchi, Y. and Shimada, H.
Malware Detection based on HTTPS Characteristic via Machine Learning.
DOI: 10.5220/0006654604100417
In Proceedings of the 4th International Conference on Information Systems Security and Privacy (ICISSP 2018), pages 410-417
ISBN: 978-989-758-282-0
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All r ights reserved
mand and control (C&C) servers. Malwares commu-
nicate through internet with a remote server that give
order to one or all the malwares that have infected
computers. C&C is especially known with botnets.
Botnets are computers that have been infected in or-
der to be controlled remotely. Generally botnet do not
do any harm to the infected computer except than sen-
ding unwanted internet traffic. Botnet are particularly
known for sending spam, creating DDos attack, etc.
Which in fact it can be harmful for the network and
also for the user as the IP address can be blacklisted
if detected for this kind of attack
1
.
There are two main reasons for malwares to have the
possibility to communicate.
First of all, the connection to a C&C server is made to
modify the behaviour of the malware for exemple de-
ciding when to launch an attack, or which data should
be stolen.
Sending the stolen data is not the only reason for mal-
wares to communicate. As malwares become more
and more efficient, they tend to behave as legitimate
programs. They can now install updates of their codes
automatically
2
. Updating the malware’s code is the
second reason of enabling malware’s communication.
Being able to update a malware give the possibility to
add new behaviour to a malware, new type of attacks
but not only, this also give the possibility to avoid their
detection. For exemple if a malware is discovered and
the antimalwares softwares start to use the signature
of the malware to detect it, it will become useless. By
updating the malware it is possible to modify its sig-
nature and make it difficult the detection of already
installed ones.
Now malwares try more and more to imitate normal
traffic that are used by everyone.
For example some botnets use Twitter
3
, Reddit or ot-
her well known web application to communicate with
the hacker.
Malware programmers can also create their own cus-
tom C&C server that do not use irc or the recent social
media.
3 RELATED WORKS
Analysing HTTP packets for security purpose was
especially done for making IDS system. Using ma-
chine learning and the use of good features and espe-
1
That would not be very nice not to be able to connect
to Google because your computer tried to DDos it.
2
Which is now done by automatically by most popular
OS.
3
https://www.welivesecurity.com/2016/08/24/first-
twitter-controlled-android-botnet-discovered/
cially algorithm is really important for creating good
detectors(Maloof, ). Often no so many features are
necessary. Previous works (Ichino et al., 2015)(Ot-
suki et al., 2014) also show that HTTP headers con-
tain a lot of information that can help for their de-
tection. This methods were base on time slot informa-
tion and the open ports for the applications. Otsuki et
al. (Otsuki et al., 2014) made his work especially on
detecting infection based on HTTP packets. In this
paper it was proven that is possible to increase mal-
ware detection, for worms and trojan malware types,
by using some of ASCII opcodes present in HTTP
packets. It also show that worms and trojan use some
HTTP client header far less than normal traffic and
that give an effective way to help discriminating this
kind of malwares from normal traffic. This method
was also link with other features such as the average
time slot interval or HTTP packet length. All of this
method are made to avoid at maximum the content
of HTTP packet, for privacy problems but also be-
cause this content can be encrypted which would not
give enough, or accurate information. However this
methods are only effective with HTTP which can be
problematic with the increase of malwares using en-
crypted communication especially with HTTPS.
The work presented in this paper is a continuation
of a paper : ”Malware Originated HTTP Traffic De-
tection” (Ogawa et al., ). This paper presents how
it is possible to detect infected computers by analy-
sing HTTP traffic network. This is possible because
now most of malwares communicate with the hacker
through C&C servers. By the help of machine lear-
ning it was proven that it is possible to detect infected
computers by analysis the HTTP network traffic ge-
nerated by a computer. This is done by looking to the
data of HTTP packet at the exception of body con-
tent, allowing kind of respect of privacy
4
. In the pre-
vious HTTPS network traffic generated by malwares
was not studied, which will be done in this paper. The
previous analyser was based on the following features
: HTTP request interval, body size, and bag-of-words
of the HTTP headers. The appearance ratio of host
pairs was then calculated before starting the classifi-
cation learning. The method was evaluated by 5-fold
cross validation and gave an average value for the re-
call of 96%. The following array show a summary
of the results. The following notation will be used
: P
Normal
for the precision in detecting normal traf-
fic and P
In f ected
for the precision in detecting infected
traffic. The same notation is used for the recall by
using R instead of P.
The protocol proposed in this paper is similar to the
one used in the previous one. The protocol will be
4
Except that there is no privacy when using HTTP
Malware Detection based on HTTPS Characteristic via Machine Learning
411
Table 1: Average for k-means with k=100.
P
N
P
I
R
N
R
I
Average 0.95 0.85 0.94 0.82
more detailed in the following section.
All this works show one common point which is that
malware detection can be done only by analysing
HTTP header, and some other features. This make
an easy way to implement the detection as it can be
made at the firewall level. However, it is only based
on HTTP and HTTPS was not studied. In this paper
will focus on HTTP and especially on HTTPS.
In June 23, 2017, Cisco published a paper about si-
milar topic : ”Detecting Encrypted Malware Traffic
(Without Decryption)”(Cis, ). The technique used dif-
ferent features from the one used in this paper. The fe-
atures used by Cisco, as described in their paper, are
the following ones :
Legacy, it correspond to classic data used to ana-
lyse network traffic packets. They used the dura-
tion of the communication, the number of packets
and number of bytes send by the client and send
by the server. It was collected by the Cisco Net-
Flow.
Sequence of packet length is also a classic fea-
ture as it correspond to the size of the packets sent
and received.
TLS Metadata is a framework that allow to re-
trieve information on TLS packets. It was used to
fingerprint the client during the checkhand proto-
col. It allowed them to determine the origin of the
certificate used for the exchange.
By using this features and a Random tree forest algo-
rithm, Cisco was able to have 99.99% of the normal
traffic correctly classified and 85.90%
5
for malware
traffic. The results of this paper will be compared to
ours in the result section.
4 TRAFFIC DETECTION
METHOD
Most of the protocol used was mostly inspired from
the method used in the previous paper. In this section
will be presented how the data is prepared from raw
pcap files, the way of analysing it and the testing met-
hod used to validate the protocol.
5
precision and recall are not precised
4.1 Testing Method
Testing a protocol is the only way to determine its
efficiency. In this section the testing method used to
validate our work is detailed.
4.1.1 Machine Learning Implementation
In order to avoid bugs or problem with machine le-
arning implementation, well known machine learning
libraries were used. In this project, the Weka
6
fra-
mework has been used. This framework developed in
Java by the University of Waikato posses a lot of ma-
chine learning algorithm already implemented. It also
has a graphical interface that allows to observe the
data very easily. For this reason this framework was
used for testing our detection method. It also allows
to use most of machine learning on data, by putting
the data into an appropriate format through a file.
4.1.2 Datasets
Two sources of data were used in order to decrease the
risk to use biased information. This two sources pro-
vided pcap files of malware network traffic or normal
network traffic. The first data source is some data pro-
vided by NTT
7
a Japanese telecommunication com-
pany. The second data source are pcap files generated
by the Georgia Institute of Technology
8
using Panda
technology to retrieve it
9
. For the normal traffic net-
work, it correspond to 12h of capture of network traf-
fic generated by our lab.
4.1.3 Efficiency Measurement
The values used to determine the pertinence of a mo-
del are recall and precision as described in the ma-
chine learning section. This values have been mea-
sured through the weka framework. On every dataset
k-fold cross-validation have been used with k equal 5
and 10 to verify that the results are the same.
4.2 Proposed Method
In this section will be explained the way of preparing
the data for the machine learning algorithm step. The
method is divided as follow:
1. Data retrieving
6
http://www.cs.waikato.ac.nz/ml/weka/
7
Nippon Telegraph and Telephone
8
panda.gtisc.gatech.edu/malrec
9
https://github.com/moyix/panda-malrec
ICISSP 2018 - 4th International Conference on Information Systems Security and Privacy
412
2. Network traffic divided by block of 30 minutes
10
3. Construction of HTTP/HTTPS request/response
pairs
4. Extraction of the features
5. Training of the classifier by using machine lear-
ning
6. Classification of traffic network
The detection method can be resumed as in Figure1
and the creation of the classifier as in Figure2.
Figure 1: Malware Detection Protocol.
Figure 2: Classifier Construction Method.
4.2.1 Construction of HTTP/HTTPS
Request/Response Pairs
The different network traffic captures are not directly
analysed. In order to extract and construct the infor-
mation Bro
11
was utilised. All the traffic captures are
divided by request/response pairs in order to isolate
communication. Request/response pairs correspond
to all the traffic exchange between two machines (i.e
10
This value of 30 minutes was choose ”randomly”. It
could be interesting to determine the optimise time (or the
optimise number of packets recived)
11
The Bro Network Security Monitor
https://www.bro.org
packet with same source and destination address and
the response to this packets). This is described in the
Figure 3.
Figure 3: request/respons pairs(Ogawa et al., ).
It is really important to separate the data of every
hosts pairs Indeed, if this is not done, we face the risk
to have to much noise due to normal traffic data
12
.
4.2.2 Creation of the Features
For HTTP/HTTPS packets different values are used
as features for training our analyser. In this part will
be explain how the data is created, and which features
are chosen for both HTTP and HTTPS.
After having isolated the different host pairs, the
fields of the packets are analysed. For every packets
the content of the different fields, except for data field,
are analysed. All of this data fields can take a finite
number of values. The different values take by the
fields during a period of time is the information that
will be used. For every possible values of a field will
be associated the frequency of appearance of this va-
lue. This frequency correspond to the number packets
having this value divided by the total amount of pac-
kets (during the fixed period of time decided before-
hand).
The fields that are analysed for HTTP and HTTPS are
the following ones :
HTTP: only the client headers are analysed, and
the frequency of appearance of each client header
is used
13
.
HTTPS: two elements characterised a HTTPS
communication: a protocol and the encryption al-
gorithm used by this protocol. We create a vec-
tor indicating the TLS version used during the ex-
change (TLS 1.0, 1.1 or 1.2), in this vector the en-
12
In your infected computer there is not only malwares
running but also your web browser, email browser...
13
And it will be shown that it is sufficient to discriminate
normal traffic to infected traffic.
Malware Detection based on HTTPS Characteristic via Machine Learning
413
cryption algorithm used is also indicated. The use
of this two informations are used to discriminate
the different HTTPS communications.
5 RESULTS
5.1 Values
In this section the values obtained for both HTTP and
HTTPS will be given. The value used are the preci-
sion and recall for both normal and infected network
traffic. The precision and recall obtain are relatively
high. To obtain this results different machine learning
were compared. We only give in this paper the best
results obtained which was done by using AdaBoost
on J48.
5.1.1 HTTP
The results obtained this time by only taking into ac-
count the content of HTTP client headers gave slig-
htly better results than the previous study. It seems
that the time interval and the other informations of
the previous study add some noise that make decrea-
ses the precision of the detection
14
.
The results show that most of the malwares can be de-
tected by this method (more than 99%) however there
is a small part of the normal network traffic (less than
7%) that is not correctly detected.
Our detection rate is similar to the most popular an-
tivirus in the market
15
however our number of false
positive is slightly higher
16
than most antivirus (vir,
). However as 7% of normal traffic is considered as
malware it can lead to an increase of threats, as it will
decrease the reliance on the detection. Decreasing the
number of false positive should be the next step to
achieve in order to improve this method. Adding a
white list for the false positives can also be an option.
Table 2: Average for HTTP using AdaBoost on J48.
P
N
P
I
R
N
R
I
Average 0.94 0.99 0.93 0.99
5.1.2 HTTPS
The results obtained for detecting malwares using
HTTPS are similar to the one for malwares using
14
that was already very high.
15
99.8% for the best one and 90% for the worst one.
16
Except that it is difficult to compare as they give a num-
ber and not a percentage
HTTP. The results were obtained by taking different
information the TLS version alone, the cipher mode
alone and TLS version and cipher mode together. The
different models give, of course, different results but
with similar values.
The model using the TLS version alone presents a
very good value (more than 0.99) for the recall of in-
fected traffic. However at the exception of true po-
sitive value they are all under 0.97 and with a false
negative rate of 0.91 it cannot be used in real environ-
ment.
Using cipher mode only gives better results than the
TLS version alone but still with a not so good value
for the false positive rate (0.95).
By combining TLS version and cipher mode we
obtain even better results than for virus using HTTP,
surprisingly we obtain a better recall than with HTTP
(0.97 against 0.93). The results are very close to the
model using cipher mode alone. The false positive
rate decreased and pass from 0.1 (for TLS vesion) to
0.03, which is more acceptable even if a white list is
still necessary to avoid at maximum this situation.
TLS version and cipher mode give by themselves a lot
of information concerning the nature of the network
traffic analysed. This can suggest that virus do not
behave as normal programmes even while communi-
cating on Internet. It will be analysed on the following
section.
Table 3: Average for HTTPS AdaBoost on J48.
P
N
P
I
R
N
R
I
TLS version 0.97 0.97 0.91 0.99
Cipher mode 0.99 0.98 0.95 0.99
TLS version &
cipher mode
0.99 0.99 0.97 0.99
5.2 Interpretation
The results detailed in the previous section show that
we can obtain good results on detecting infected com-
puter without having so much information on the sent
packet
17
. That means the information given by every
element is high. We analysed all this information to
see how easy or not it is to discriminate infected ma-
chines from non infected ones. In order to obtain the
information we look on the average of the average ap-
pearance of every elements for every host pairs. This
statistics show that some values are a strong indicator
of infection or not of a machine. This indicators will
be analysed for both the HTTP and HTTPS study.
17
Two elements for HTTPS.
ICISSP 2018 - 4th International Conference on Information Systems Security and Privacy
414
Figure 4: Average of client header for HTTP.
5.2.1 HTTP
As it can be observed on Figure4 representing the
average appearance of client header for HTTP reque-
sts, normal and infected machines do not use so often
every header.
The interesting headers are not the ones presen-
ting the same values but header that are not used at
the same way. Some of the headers that are not use
as the same way by malwares will be listed and the
interpretation given.
CONNECTION: is a control option for the cur-
rent connection. It should not be used for HTTP2.
It is used more than 95% of the time by normal
communication and more than 50% by malwares
generated communication. The reason is that mal-
wares used both HTTP1 and HTTP2 while most
websites are still using HTTP1
18
ACCEPT, ACCEPT-ENCODING, ACCEPT-
LANGUAGE: are used for defining the accepted
contents and encoding for the response. It is used
less than half of the time in malwares communica-
tion compare to normal ones. The accepted con-
tents or encoding are generally known beforehand
by the hacker. Making it not useful to use.
COOKIE: contain the value of a cookie previ-
ously sent by the server. Used 30% of the time
in normal traffic and less than 10% of the time
by malwares. This is probably due to the reason
that most of malwares only connect to a server and
there is no need of recognising the user (that is the
infected machine)
19
CACHE-CONTROLS: is used to control how
the cache system works, and especially if the ca-
che should be used for the packets received. With
18
Less than 13% of the most visited websites use HTTP2.
https://w3techs.com/technologies/details/ce-http2/all/all
19
Or made by another way than sending cookie
the data provided it seems that malwares tend to
use more the possibilties of controling the ca-
che with 70% of the packets received, than nor-
mal communications which reach only 35% of the
packets.
IF-MODIFIED-SINCE: a response is sent only
if a modification happened. This is generally used
to avoid loading content if no modification occur-
red, especially when there is a cache system. It is
mostly not used by malwares.
5.2.2 HTTPS
Similar analysis than the one made on HTTP can be
done on HTTPS. It will refer to Figure5 and Figure6.
First concerning the cipher mode used, for normal
communication AES128 implementation using SHA-
2 or SHA-3 are used. However for malware commu-
nication, a RC4 implementation using MD5 is used.
This is very surprising when it is well known that
MD5 is deprecated and no cryptographic security is
given. Approximately 40% of the packets send in
HTTPS by malwares are using cryptographic imple-
mentation using MD5. Blocking such packet could
lead to suppress a big part of the malware traffic as it
mostly not used in normal traffic. Using the exploit
on such deprecated protocol could also be a possibi-
lity to discriminate malware traffic from normal one.
This should be verified in another study.
Also, it happens that the cipher mode is not commu-
nicate or unknown
2021
, this is probably because hac-
kers already know which algorithm will be used. But
not giving this information gives a lot of information
about the process sending this packets. As no normal
traffic packet as been send with unknow field for the
20
Or at list Bro IDS was not able to recognise the cipher
mode
21
not shown on the Figure
Malware Detection based on HTTPS Characteristic via Machine Learning
415
Figure 5: Average of cipher mode for HTTPS.
Figure 6: Average of TLS version for HTTPS.
cipher mode used, it gives a good way of detecting
this kind of malwares.
A similar behaviour as been observed with the TLS
version. Some of malwares do not indicate the TLS
version used
22
. Giving the possibility to detect simi-
lar malwares. This detection can be done by indica-
ting directly as suspect process sending packet and not
precise the cipher mode or TLS version. The reason
of this field not indicated should be study. Maybe this
malware prentends to use TLS or use its own one .
22
And even sometimes modifying the value during the
communication.
5.2.3 Comparison with Cisco Results
As Cisco paper presents some techniques similar
to ours, we will compare them. First of all Cisco
paper present a very good detection of the nor-
mal files (near 100% and reach it under certain
parameters but with high rate of false positive)
which is far more better than our method as we
have approximately 7% of false positive. For the
detection of malwares the results is similar to our
method but with different parameters. When they
reach 100% for the normal traffic detection rate
the malware traffic detection is low. However
by decreasing this detection we obtain similar
results for malware detection however they obtain
ICISSP 2018 - 4th International Conference on Information Systems Security and Privacy
416
a better normal traffic detection
23
. By using two dif-
ferent methods we obtain similar results (with better
precison in Cisco paper). This is probably due to the
information added by TLS metadata.
One possibility of further work would be to extend
our method by using TLS metadata in order to quan-
tify the information given by it. It would be also in-
teresting to test both method with the two datasets :
ours and Cisco’s one.
5.3 Malwares Do Not Respect Rules
The previous section show that malwares do not be-
have in the same way than normal process.
For HTTP traffics the header used are not the same
and for headers shared by both type of traffic, the
average frequency of using the headers is comple-
tely different. Normally headers are used to give in-
formation to web sites in order to make them adapt
correctly the content for the client. But malwares
know in advance, at least potentially, the characte-
ristics of the C&C server they connect to. Malwa-
res when they communicate with custom C&C server,
do not communicate with website, hackers do not use
this kind of headers probably because hackers alre-
ady know, when programming malwares, which and
how the information will be send. It is the same for
HTTPS when TLS version or cipher mode are not pre-
cise. However for HTTPS it is even more surprising
as malwares use mostly deprecated protocols or do
not precise the encryption method and version as it is
mandatory contrary to some HTTP headers. This is
really easy to detect while observing network traffic
packets. Because of that, an easy way to protect a sy-
stem is to send and alert when deprecated protocol are
used
24
and set a white list for old programs that can-
not be changed and that use this kind of deprecated,
and unsafe protocols.
However, this elements that can be used to detect mal-
wares based on traffic analysis can be modified, es-
pecially for HTTPS. If hacker modified the way of
constructing malwares by using classic cipher mode
and TLS version for HTTPS and using more classic
HTTP header, it will probably be necessary to use ot-
her features to detect malwares by this method. The
following step would be to find features not based on
the cipher mode used or easy modifiable header va-
lues and that describe more precisely the behaviour
of malwares.
23
But difficult to know exactly as we do not know what
is the value, probably precision only, and the recall is not
precised.
24
As using deprecated protocol for cryptography it is re-
ally not recommended
6 CONCLUSION
Infected machine based on the analysis of the net-
work traffic generated is relatively efficient. With
more than 90% of precision and recall for both HTTP
and HTTPS, with only few features. This is due to
the fact that malwares are made to be only malwares
and hackers already fixed how the data will be sent
both by the malware and the C&C server. And also
for some unknow reason they use deprecated proto-
col for SSL communication. However a modification
of the packet sent could lead to the impossibility of
detection by this method. There are different possi-
bilities for future work. The first one is to find anot-
her features in the case of modification of the network
behaviour by the hacker as explained before. The se-
cond is link to real environment implementation. The
use for real time detection was not tested during this
work. It would be of great use to see if it can be used
in real time
25
and especially if it can detect malware
that were not detected by classical antivirus. This
method should be used complementary to other pro-
tection method. For further research, an analysis of
the cipher suite used and the reason of why the data
of TLS cipher suite and TLS version is not given du-
ring malware communication should be study.
REFERENCES
http://www.av-comparatives.org/wp-content/uploads/2014/
04/avc fdt 201403 en.pdf.
Detecting encrypted malware traffic (without decryp-
tion). https://blogs.cisco.com/security/detecting-
encrypted-malware-traffic-without-decryption.
Ichino, M., Kawamoto, K., Iwano, T., Hatada, M., and Yos-
hiura, H. (2015). Evaluating header information fea-
tures for malware infection detection. Journal of In-
formation Processing, 23(5):603–612.
Maloof, M. A. Machine Learning and Data Mining for
Computer Security : Methods and Application. Lon-
don Spring.
Nasi, E. (2014). Bypass antivirus dynamic analysis. http://
packetstorm.foofus.com/papers/virus/BypassAV Dy-
namics.pdf.
Ogawa, H., Yamaguchi, Y., Shimada, H., Takakura, H.,
Akiyama, M., and Yagi, T. Malware originated
http traffic detection utilizing cluster appearance ratio.
ICOIN 2017.
Otsuki, Y., Ichino, M., Kimura, S., Hatada, M., and Yoshi-
ura, H. (2014). Evaluating payload features for mal-
ware infection detection. Journal of Information Pro-
cessing, 22(2):376–387.
Page, C. R. E. (2003). Anti-debugging & software pro-
tection advice.
25
and real environment not virtual machine
Malware Detection based on HTTPS Characteristic via Machine Learning
417