The Curious Case of Machine Learning in Malware Detection

Sherif Saad1 , William Briguglio1 and Haytham Elmiligi2

1 Schoolof Computer Science, Windsor University, Canada
2 Computing Science Department, Thompson Rivers University, Canada
{shsaad, briguglwb}@uwindsor.ca, [email protected]

Keywords: Malware, Machine Learning, Behaviour Analysis, Adversarial Malware, Online Training, Detector Interpre-
tation.

Abstract: In this paper, we argue that detecting malware attacks in the wild is a unique challenge for machine learn-
ing techniques. Given the current trend in malware development and the increase of unconventional malware
attacks, we expect that dynamic malware analysis is the future for antimalware detection and prevention sys-
tems. A comprehensive review of machine learning for malware detection is presented. Then, we discuss how
malware detection in the wild present unique challenges for the current state-of-the-art machine learning tech-
niques. We defined three critical problems that limit the success of malware detectors powered by machine
learning in the wild. Next, we discuss possible solutions to these challenges and present the requirements of
next-generation malware detection. Finally, we outline potential research directions in machine learning for
malware detection.

1 INTRODUCTION vioral malware analysis will dominate the next-

generation malware detection systems.
Nowadays, computer networks and the Internet have There is a general belief among cybersecurity ex-
become the primary tool for spreading and distribut- perts that antimalware tools and systems powered by
ing malware by malware authors. The massive num- artificial intelligence and machine learning will be the
ber of feature-rich programming languages and off- solution to modern malware attacks. The number of
the-shelf software libraries enable the development of studies published in the last few years on malware
new sophisticated malware such as botnet, fileless, detection techniques that leverage machine learning
k-ary and ransomware. New computing paradigms, is a distinct evidence of this belief as shown in sec-
such as cloud computing and the Internet of Things, tion 2. In the literature, various malware detection
expand potential malware infection sites from PC’s to techniques using machine learning are proposed with
any electronic device. excellent detection accuracy. However, malware at-
To decide if software code is malicious or benign tacks in the wild continue to grow and manage to
we could either use static analysis or dynamic analy- bypass malware detection systems powered by ma-
sis. Static analysis techniques do not execute the code chinelearning techniques. This because it is difficult
and only examine the code structure and other binary to operate and deploy machine learning for malware
data properties. Dynamic analysis techniques, on the detection in a production environment or the perfor-
other hand, execute the code to observe the execution mance in a production environment is disturbing (e.g.,
behaviors of the code over the network or at endpoint high false positives rate). In fact, there is a signif-
devices. Some malware detection systems apply only icant difference (a detection gap) between the accu-
static or dynamic techniques, and some apply both. racy of malware detection techniques in the literature
While dynamic malware analysis techniques are not and their accuracy in a production environment.
intended to replace static analysis techniques, recent A perfect malware detection system will detect all
unconventional malware attacks (botnet, ransomware, types of malicious software and will never consider a
fileless, etc) and the use of sophisticated evasion tech- benign software as a malicious one. Cohen provided a
niques to avoid detection have shown the urgent need formal proof that creating a perfect malware detection
of dynamic analysis and the limitations of static anal- system is not possible (Cohen, 1987; Cohen, 1989).
ysis. In our opinion, the use of dynamic and beha- Moreover, Chess and White proved that a malware
detector with zero false positives is not possible (Raff et al., 2017). The proposed technique achieved
(M. Chess and R. White, 2000). Selcuk et al. dis- 94.0% detection accuracy.
cussed the undecidable problems in malware detec- Several works have been proposed to detect An-
tion in more details (Selcuk et al., 2017). In light of droid malware apps using static analysis techniques.
this, the high levels of accuracy claimed by commer- Sahin et al. proposed an Android malware detection
cial malware detection systems and some malware de- model that uses app permission to detect malicious
tection studies in literature seems questionable. apps (OSahn et al., 2018). They used the permissions
In this paper, we briefly review the current state required by the app with a weighted distance function
of the art in malware detection using machine learn- and KNN and Naive Bayes classifier to detect mali-
ing approach. Then, we discuss the importance of cious apps. They reported an accuracy up to 93.27%.
dynamic and behavioral analysis based on emerg- Su and Fung used sensitive functions and app permis-
ing malware threats. Next, the shortcomings of the sions to detect Android malware (Su and Fung, 2016).
current machine learning malware detectors are ex- They used different machine learning algorithms such
plained to indicate their limitations in the wild. Fi- as SVM, decision tree, and KNN to build an android
nally, we discuss the possible solutions to improve malware detector. They reported an average accuracy
the quality of malware detection systems and point between 85.0% and 90.0%
out potential research directions. Collecting and monitoring all malware behaviors
is a complicated and time-consuming process. For
that reason, several works in the literature focused
2 LITERATURE REVIEW on collecting partial dynamic behaviors of the mal-
ware. Lim et al. (Lim et al., 2015) proposed a mal-
ware detection technique by analyzing network traf-
In recent years, machine learning algorithms have
fic generated when the malware communicates with
been used to design both static and dynamic analy-
a malicious C&C server such as in the case of bot-
sis techniques for malware detection. Hassen et al.
net or ransomware. The proposed technique extracts
proposed a new technique for malware classification
a set of features from network flows to present a flows
using static analysis based on control statement shin-
sequence. The authors used different sequence align-
gling (Hassen et al., 2017). In their work, they used
ment algorithms to classify malware traffic. They re-
static analysis to classify malware instance into new
ported an accuracy above 60% when analyzing mal-
or known malware families. They extracted features
ware traffic in a real network environment.
from disassembled malicious binaries and used ran-
dom forest algorithm to classify malware using the Kilgallon et al. applied machine learning and dy-
extracted features. Using a dataset of 10,260 malware namic malware analysis (Kilgallon et al., 2017). The
instances, they reported up to 99.21% accuracy. proposed technique gathers register value information
Static analysis has been used to study malwares and API calls made by the monitored malware bi-
that infect embedded systems, mobile devices, and naries. The collected information is stored in vec-
other IoT devices. Naeem et al. proposed a static tor structures and analyzed using a value set analysis
analysis technique to detect IoT malware (Naeem method. Then, they used a linear similarity metric to
et al., 2018). The proposed technique converts a mal- compare unseen malware to known malware binaries.
ware file to a grayscale image and extracts a set of Their experiment showed that the proposed technique
visual features from the malware image to train an could detect malware with an accuracy up to 98.0%
SVM classifier that could distinguish between mal- Omind and Nathan proposed a behavioral-based
ware families using visual features. Using a dataset malware detection method using a deep belief net-
of 9342 samples that belong to 25 malware families, work (David and Netanyahu, 2015). The proposed
they reported 97.4% accuracy. Su et al. proposed a method collected data about malware behaviors from
similar technique to classify IoT malware into mal- a sandbox environment. The collected data is API
ware families using visual features and image recog- calls, registry entries, visited websites, accessed ports,
nition (Su et al., 2018). Their approach is very simi- and IP addresses. Then using a deep neural net-
lar to the one proposed in (Naeem et al., 2018). They work of eight layers, it generates malware signatures.
used a one-class SVM classifier and tested their ap- These signatures could be used to train malware de-
proach on IoT malwares that infect Linux-like IoT tectors. In their experiments, they reported up to
systems; they reported 94.0% accuracy for detect- 95.3% detection accuracy with a malware detector
ing malware and 81.8% accuracy for detecting mal- utilizing the SVM algorithm.
ware families. Raff et al. proposed a malware detec- Yeo et al. proposed a new malware detection
tion technique using static analysis and deep learning method by monitoring malicious behaviors in net-
work traffic (Yeo et al., 2018). They designed 35 detectable. Here we describe several recent examples
features to describe malicious traffic of malware in- of new malware threats that are difficult to detect or
stances. They tested several machine learning al- analyze using static analysis.
gorithms including CNN, MLP, SVM, and random For instance, the Internet of Things (IoT) is an ap-
forest. The proposed method achieved an accuracy pealing platform for modern and sophisticated mal-
above 85% when utilizing CNN or random forest. ware scuh as ransomware. Zhang-Kennedy et al. dis-
Prokofiev et al. proposed a machine-learning tech- cussed the ransomware threat in IoT and how a self-
nique to detect C&C traffic of infected IoT devices spreading ransomware could infect an IoT ecosystem
(Prokofiev et al., 2018). The proposed approach used (Zhang-Kennedy et al., 2018). The authors pointed
network traffic features such as port number, IP ad- out that the ransomware will mainly lock down IoT
dresses, connection duration and frequency. They re- devices and disable the essential functions of these
ported a detection accuracy up to 97.3%. However, devices. The study focused on identifying the attack
the proposed approach is still relying on traditional vectors in IoT, the techniques for ransomware self-
malware analysis methods and will not be able to spreading in IoT, and predicting the most likely class
work in production IoT deployment as discussed in of IoT applications to be a target for ransomware at-
(Soliman et al., 2017). Several hybrid malware de- tacks. Finally, the authors identified the techniques
tection techniques that combine both static and dy- the ransomware could apply to lock down IoT de-
namic analysis have also been proposed (Martinelli vices. Authors in (Zhang-Kennedy et al., 2018) used
et al., 2016; Paola et al., 2018). These techniques try a Raspberry to develop a proof of concept IoT ran-
to improve the quality and performance of malware somware that can infect an IoT system. One inter-
detection systems by taking advantage of static and esting aspect in (Zhang-Kennedy et al., 2018) is the
dynamic analysis to build robust malware detection need for collaboration or swarming behavior in IoT
systems. ransomware, where the IoT ransomware will spread
as much as possible and then lock down the devices
or lock down the device and then spread.
3 EMERGING MALWARE Miller and Valasek developed a proof-of-concept
THREATS for malicious code that infects connected cars and
lockdown key functions (Miller and Valasek, 2015).
For instance, the authors demonstrated the ability for
With the recent changes in malware development
the malicious code to control the steering wheel of a
and the rise of commercial malware (malicious code
vehicle, disable the breaks, lock doors, and shut down
rented or purchased), many new challenges are facing
the engine while in motion. Behaving as ransomware,
malware analysts that make static analysis more diffi-
this real example of a malware that locks and disables
cult and impractical. These challenges will force an-
key features in IoT systems (e.g. connected cars)
timalware vendors to adapt behavioral malware anal-
could have life threatening consequences if the ran-
ysis and detection techniques. In our opinion, there
som is not paid. The study explained a design flow
are two main reasons behind these challenges; the rise
in the Controller Area Network (CAN) protocol that
of unconventional computing paradigms and uncon-
allows malicious and crafted CAN messages to be in-
ventional evasion techniques. There is a new gener-
jected into the vehicle CAN channel by a compro-
ation of malwares that take advantage of unconven-
mised mobile phone that is connected to the vehicle
tional computing paradigms and off-the-shelf soft-
entertainment unit. It was reported that for some ve-
ware libraries written by feature-rich programming
hicles only the dealership could restore and patch the
languages. The current state-of-the-art malware anal-
vehicle to prevent this attack. Choi et al. proposed a
ysis/detection techniques and tools are not effective
solution for malware attacks in connected vehicles us-
against this new generation of malware.
ing machine learning (Choi et al., 2018). The solution
uses SVM to distinguish between crafted malicious
3.1 Unconventional Computing CAN messages, and benign CAN messages generated
Paradigms by actual electronic control units (ECU). The model
extracts features from the vehicle ECUs and creates
New computing paradigms and technologies such as fingerprints for those ECUs. The ECU fingerprint is
cloud computing, the internet of things, big data, in- noticeable in a benign CAN message and does not ex-
memory computing, and blockchain introduced new ist in a malicious message
playgrounds for malware authors to develop com- Azmoodeh et al. discussed a new technique to de-
plex and sophisticated malwares that are almost un- tect ransomware attacks in IoT systems by monitor-
ing the energy consumption of infected devices (Az- According to KASPERSKY lab, 140 enterprises were
moodeh et al., 2018). As a proof of concept, they attacked in 2017 using fileless malwares (Global Re-
studied the energy consumption of infected Android search and Analysis Team, KASPERSKY Lab, 2017).
devices. The devices were infected by a ransomware Ponemon Institute reported that 77% of the attacks
with crypto impact. They used different machine against companies use fileless techniques (Ponemon
learning models (KNN, SVM, NN, and Random For- Institute, Research Report, 2017). Moreover, there
est) to analyze energy consumption data and extract are several signs that ransomware attacks are going
unique patterns to detect compromised Android de- fileless, as discussed in (Magnusardottir, 2018). Be-
vices. They reported a ransomware detection accu- sides these signs, there are other reasons in our opin-
racy of 95.65%. ion that confirms that ransomware and other malware
In 2015, Karam (INTERPOL) and Kamluk attacks will be fileless. One main reason is the mov-
(Kaspersky lab) introduced a proof of concept ing towards in-memory computing.
distributed malware that also takes advantage of In recent years, in-memory computing and in-
blockchain technology (Karam and Kamluk, 2015). memory data stores became the first backbone and
In 2018, Moubarak and et al. provided design and storage technology for many organizations. Many big
implementation of a K-ary malware (distributed mal- data platforms and data grids (Apache Spark, Redis,
ware) that takes advantages of the blockchain net- HazelCast, etc.) enable storing data in memory for
works such as Etherum and Hyperledger (Moubarak performance and scalability requirements. Valuable
et al., 2018). The proposed malware is stored and ex- data and information is stored in memory for a long
ecuted inside blockchain networks and acts as a ma- time before moving to a persistent data store. In-
licious keylogger. While detecting a K-ary malware Memory ransomware that encrypts in-memory data
is an NP-hard problem(de Drézigué et al., 2006), it (such as recent transactions, financial information,
is also complicated to implement a K-ary malware. etc.) present a severe and aggressive attack. This is
However, Mubarak’s works demonstrated the sim- because any attempt to reset or report the machine
plicity of K-ary malware development by taking ad- to remove the ransomware from the device memory
vantage of blockchain technology as distributed and or shutdown the application will result in losing this
decentralized network. valuable data permanently.
The moving towards distributed and decentralized
3.2 Unconventional Evasion Techniques computing is another reason for the rise of fileless ran-
somware. In distributed and decentralized computing
The new generation of malware will use advanced several nodes and devices are available to store the
evasion techniques to avoid detection by antimalware in-memory malware, which will increase the life ex-
systems and tools. New evasion techniques imple- pectancy of the malware since there will always be a
mented by malware authors use new technologies and group of active nodes were the malware could repli-
off-the-shelf software libraries that enable the design cate and store itself.
of sophisticated evasion methods. Antimalware ven- The recent and massive development in machine
dors and malware researchers discussed recent exam- learning /artificial intelligence (aka data science) and
ples of using new antimalware evasion techniques in a large number of off-the-shelf machine learning li-
the wild. braries enable malware authors to develop advanced
Fileless malware or memory-resident malware is evasion techniques.
the new technique used by malware authors to de- Rigaki and Garcia proposed the use of deep learn-
velop and execute malicious attacks. Fileless mal- ing techniques to create malicious malware sam-
ware resides in device memory and does not leave any ples that evade detection by mimicking the behav-
files on the infected device file system. This makes iors of benign applications (Rigaki and Garcia, 2018).
the detection of the fileless malware using signature- In their work, a proof of concept was proposed to
based detection or static analysis infeasible. In addi- demonstrate how malware authors could cover the
tion, the fileless malware takes advantage of the utili- malware C&C traffic. The authors use a Genera-
ties and libraries that already exist in the platform of tive Adversarial Networks (GANs) to enable malware
the infected device to complete its malicious intents. (e.g., botnet) to mimic the traffic of a legitimate ap-
In other words, benign applications and software li- plication and avoid detection. The study showed that
braries are manipulated by fileless malware to accom- it is possible to modify the source code of malware
plish the attack objectives. to receive parameters from a GAN to change the be-
Fileless malware attacks and incidents are already haviors of its C&C traffic to mimic the behaviors of
observed in the wild compromising large enterprises. other legitimate network applications, such as Face-
book traffic. The enhanced malware samples were cessing, and e-commerce. Malware instances evolve
tested against the Stratosphere Linux IPS (slips) sys- and change their behaviors over a short period; some
tem, which uses machine learning to detect malicious studies by antimalware vendors reported that a new
traffic. The experiment showed that 63.42% of the malware instance could change its behaviors in less
malicious traffic was able to bypass the detection. than 24 hours since it has been released (Gupta et al.,
A research team from IBM demonstrated the use 2009; Allix et al., 2015). This means a frequently
of artificial intelligence to engineering malware at- trained machine learning model will become out-
tacks (Kirat et al., 2018). In their study, the authors dated. This also means we need to frequently retrain
proposed DeepLocker as a proof of concept to show our malware detectors to be able to detect new and
how next-generation malware could leverage artificial mutated malware instances. Therefore, adaptability
intelligence. DeepLocker is a malware generation en- in machine learning models for malware detection is
gine that malware author could use to empower tra- a crucial requirement and not just a ancillary capabil-
ditional malware samples such as WannaCry with ar- ity.
tificial intelligence. A deep convolutional neural net- Recently, the challenge of adaptability, and scal-
work (CNN) was used to customize a malware attack ability of machine learning models for malware de-
by combining a benign application and a malware tection in the wild has become obvious (Narayanan
sample to generate a hybrid malware that bypasses et al., 2016). The majority of the work proposed in
detection by exposing (mimicking) benign behaviors. the literature have done very little to reduce and op-
Besides that, the malware is engineered to unlock its timize the feature space to design detectors ready for
malicious payload when it reaches a target (endpoint) early malware detection in a production environment
with a loose predefined set of attributes. In the study, (Hajmasan et al., 2017). For instance, it is not clear
those attributes were the biometrics feature of the tar- how the proposed detection methods will scale when
get such as facial and voice features. The malware the number of monitored endpoints increase. Un-
uses CNN to detect and confirm target identity, and like computer vision, natural language processing and
upon target confirmation, an encryption key is gen- other areas that utilize machine learning, malware in-
erated and used by the WannCry malware to encrypt stances continue to evolve and change. This mostly
the files on the target endpoint device. The encryption requires retraining machine learning models in pro-
key is only generated by matching the voice and the duction, which is an expensive and complicated task.
facial features of the target. This means reverse engi- Therefore, when using machine learning for malware
neering the malware using static analysis is not useful detection, we need to think differently. New methods
to recover the encryption key. to reduce the cost of retraining malware detectors and
improve the detection quality are urgent.

4 PRACTICAL CHALLENGES 4.2 Malware Detector Interpretability

The new and emerging malware threats discussed in Cybersecurity analysts always prefer solutions that
section 3 provide strong evidence for the need of are interpretable and understandable, such as rule-
adopting dynamic and behavioral analysis to build based or signature-based detection. This is because
malware detection tools. The use of machine learning of the need to tune and optimize these solutions to
is the most promising technique to implement mal- mitigate and control the effect of false positives and
ware detectors and tools that apply behavioral analy- false negatives. Interpreting machine learning mod-
sis as shown in section 2. While the use of machine els is a new and open challenge (Shirataki and Ya-
learning for malware detection has shown promising maguchi, 2017). However, it is expected that an in-
results in both static and dynamic analysis, there are terpretable machine learning solution will be domain
significant challenges that limit the success of ma- specific, for instance, interpretable solutions for ma-
chine learning based malware detectors in the wild. chine learning models in healthcare are different than
solutions in malware detection (Ahmad et al., 2018).
4.1 Cost of Training Detectors Any malware detector will generate false posi-
tives, and unless malware analysts can understand and
The first challenge is the cost of training and updating interpret the reason that a benign application wrongly
malware detectors in production environment. Mal- classified as malicious, they will not accept those
ware detection is unlike other domains where ma- black box malware detectors. To our knowledge, no
chine learning techniques have been applied success- work in the literature investigated the interpretability
fully such as computer vision, natural language pro- of machine learning models for malware detection.
4.3 Adversarial Malware is built to detect the behaviors of a specific malware
instance (e.g., Mirai, WannaCry), or at most a sin-
Last but not least, a malware detection system utiliz- gle malware family (a group of similar malware in-
ing machine learning could be defeated (bypassed) stances). Also, each model or detector is built us-
using adversarial malware samples. For instance, ing features that are similar, such as having the same
Kolosnjaji et al. showed in (Kolosnjaji et al., 2018) computational cost, or unique to the specific execu-
that by using an intelligent evasion attack they can tion environment. This is because out of the superset
defeat the deep learning detection system proposed in of features designed to detect malware, it is common
(Raff et al., 2017) by Raff et al. They simply used that a subset of these features could be more or less
their knowledge of how the proposed deep learning useful to detect a specific malware instance or fam-
detection system operates and designed a gradient- ily. The use of micro (small) and focused detectors
based attack as an evasion technique to overcome it. reduce the cost of retraining and deployment in pro-
With adversarial malware, the system detection accu- duction. This is because detectors for new malware
racy dropped from 94.0% to almost 50.0%. Machine could be trained and added without the need to retrain
learning algorithms are not designed to work with ad- existing detectors. In addition, when a malware de-
versarial examples. Grosse et al. demonstrated that tector becames outdated as a result of malware evolv-
using adversarial malware samples; they could reduce ing behaviors, the outdated detectors are disposed of
the detection accuracy of a malware detection sys- and replaced by new ones. The use of micro-detectors
tem that uses static analysis and machine learning to enables adaptability by design rather than attempting
63.0% (Grosse et al., 2017). They also showed that to change machine learning models and algorithms to
adopting anti adversarial machine learning techniques support adaptability.
used in computer vision is not effective in malware
detection. Yang et al. proposed adversarial training as 5.2 Analyst Friendly Interpretation
a solution for adversarial malware (Yang et al., 2017).
They designed a method for adversarial android mal-
ware instances generation. The proposed method re- Adopting sophisticated machine learning techniques
quires access to the malware binaries and source code, for malware detection in a production environment is
besides, it is mainly useful for static malware detec- a challenge. This is because most of the time it is not
tion systems. possible to understand how the machine learning sys-
tems make their malware detection decisions. There-
fore, tuning and maintaining these systems is a chal-
lenge in production and new techniques for malware
5 BRIDGING THE DETECTION analysts to interpret and evaluate the performance of
GAP malware detectors are needed. We propose the use
of evolutionary computation techniques such as ge-
To overcome the challenges we discussed in section 4, netic algorithms or clonal selection algorithms to gen-
we propose new solutions to mitigate these challenges erate an interpretation for black-box machine learn-
and reduce the gap. ing models such as deep learning. Using evolutionary
computation, we could describe the decisions of mal-
5.1 Disposable Micro Detectors ware detectors using a set of IF-Then rules. The only
information required is the input features the malware
Current best practices in constructing and build- detector uses to make a decision.
ing machine learning models follow a mono- The IF-Then rules are useful to explain the behav-
lithic architecture. In monolithic architecture, iors that trigger a specific decision (e.g., malicious
a computationally-expensive single-monolithic (to or benign) by the malware detector. Cybersecurity
build and train) machine learning model is used to and malware analyst are comfortable working with
detect malwares. While this architecture or approach IF-Then rules. These rules will help in understand-
for building machine learning models is successful in ing the decision made by malware detectors, explain
other domains, we believe it is unsuitable for mal- the scope of the detection, and identify potential over
ware detection given the highly evolving character- generalization or overfitting that could result in false
istics of malware instance. We propose a new ap- positives or false negatives.
proach inspired by microservices architecture. In this It is essential that the IF-Then rules set interpre-
approach, multiple, small, inexpensive, focused ma- tation of the malware detector to be expressed in raw
chine learning models are built and orchestrated to malware behaviors and not in machine learning fea-
detect malware instances. Each model or detector tures. Machine learning features are most likely un-
derstandable by machine learning engineers and ex- 6 CONCLUSION
perts. The interpretation should be acceptable to a
malware analyst who does not need to be machine In this paper, we reviewed the current state-of-the-art
learning experts. in malware detection using machine learning. We dis-
cussed the recent trends in malware development and
5.3 Anti Adversarial Malware emerging malware threats. We argued that behavioral
analysis would dominate the next generation antimal-
To improve the resilience of malware detectors ware systems. We discussed the challenges of apply-
against adversarial malware, we believe it is essential ing machine learning to detect malware in the wild
to study the effort required by the malware authors and proposed our thoughts on how we could over-
to design an adversarial malware for specific malware come these challenges. Machine learning malware
detectors. For example, what technique a malware au- detectors require inexpensive training methods; they
thor would use to probe and study a malware detector need to be interpretable for the malware analysts and
in production to design a malware that could bypass not only for machine learning experts. Finally, they
this detector. need to tolerate adversarial malware by design.
Measuring the effort to probe detectors and design
adversarial malware under two main settings is essen-
tial. The first setting is black-box, where the malware REFERENCES
authors have minimum knowledge about the malware
detector internal design and the features used by the Ahmad, M. A., Teredesai, A., and Eckert, C. (2018). In-
machine learning algorithm. The second setting is terpretable machine learning in healthcare. In 2018
white-box, where the malware authors have sufficient IEEE International Conference on Healthcare Infor-
knowledge about the malware detector internal design matics (ICHI), pages 447–447.
and the machine learning algorithm. Training and up- Allix, K., Bissyandé, T. F., Klein, J., and Le Traon, Y.
dating the malware detectors is likely the most effi- (2015). Are your training datasets yet relevant? In
Piessens, F., Caballero, J., and Bielova, N., editors,
cient solution against adversarial malware. Knowing Engineering Secure Software and Systems, pages 51–
the effort needed to evade a malware detector will 67, Cham. Springer International Publishing.
help in designing training strategies and policies to Azmoodeh, A., Dehghantanha, A., Conti, M., and Choo,
increase the effort required to evade the detectors. K.-K. R. (2018). Detecting crypto-ransomware in
As we mentioned before, Cohen provided a for- iot networks based on energy consumption footprint.
mal proof that creating a perfect malware detection Journal of Ambient Intelligence and Humanized Com-
system is not possible (Cohen, 1987; Cohen, 1989). puting, 9(4):1141–1152.
We believe that designing a perfect adversarial mal- Choi, W., Joo, K., Jo, H. J., Park, M. C., and Lee, D. H.
ware is not possible. Therefore we expect that using (2018). Voltageids: Low-level communication char-
acteristics for automotive intrusion detection system.
ensemble-based hybrid machine learning approach IEEE Transactions on Information Forensics and Se-
for malware detector will be effective against ad- curity, 13(8):2114–2129.
versarial malware. It is expected that by creating a Cohen, F. (1987). Computer viruses: Theory and experi-
malware detector using an ensemble hybrid machine- ments. Computers & Security, 6(1):22 – 35.
learning approach, the risk of evading detection will Cohen, F. (1989). Computational aspects of computer
decrease and the effort to design adversarial malware viruses. Computers & Security, 8(4):297 – 298.
will increase. A hybrid machine learning model is David, O. E. and Netanyahu, N. S. (2015). Deepsign: Deep
when two or more different machine learning algo- learning for automatic malware signature generation
rithms are used to construct the model. In the litera- and classification. In 2015 International Joint Confer-
ture, adversarial malware samples evade malware de- ence on Neural Networks (IJCNN), pages 1–8.
tectors that use a single machine learning algorithm de Drézigué, D., Fizaine, J.-P., and Hansma, N. (2006). In-
or technique (Yang et al., 2017; Grosse et al., 2017; depth analysis of the viral threats with openoffice.org
documents. Journal in Computer Virology, 2(3):187–
Kolosnjaji et al., 2018). In our method, a hybrid ma- 210.
chine learning approach for building a malware de- Global Research and Analysis Team, KASPERSKY Lab
tector is an approach to provide a defense-in-depth (2017). Fileless attack against enterprise network.
model for malware detectors. Grosse, K., Papernot, N., Manoharan, P., Backes, M.,
and McDaniel, P. (2017). Adversarial examples for
malware detection. In Foley, S. N., Gollmann, D.,
and Snekkenes, E., editors, Computer Security – ES-
ORICS 2017, pages 62–79, Cham. Springer Interna-
tional Publishing.
Gupta, A., Kuppili, P., Akella, A., and Barford, P. (2009). Paola, A. D., Gaglio, S., Re, G. L., and Morana, M. (2018).
An empirical study of malware evolution. In 2009 A hybrid system for malware detection on big data. In
First International Communication Systems and Net- IEEE INFOCOM 2018 - IEEE Conference on Com-
works and Workshops, pages 1–10. puter Communications Workshops (INFOCOM WK-
Hajmasan, G., Mondoc, A., and Cre, O. (2017). Dynamic SHPS), pages 45–50.
behavior evaluation for malware detection. In 2017 Ponemon Institute, Research Report (2017). The 2017 state
5th International Symposium on Digital Forensic and of endpoint security risk.
Security (ISDFS), pages 1–6. Prokofiev, A. O., Smirnova, Y. S., and Surov, V. A. (2018).
Hassen, M., Carvalho, M. M., and Chan, P. K. (2017). Mal- A method to detect internet of things botnets. In
ware classification using static analysis based features. 2018 IEEE Conference of Russian Young Researchers
In 2017 IEEE Symposium Series on Computational in Electrical and Electronic Engineering (EIConRus),
Intelligence (SSCI), pages 1–7. pages 105–108.
Karam, C. and Kamluk, V. (2015). Blockchainware - de- Raff, E., Barker, J., Sylvester, J., Brandon, R., Catanzaro,
centralized malware on the blockchain. In Black Hat B., and Nicholas, C. K. (2017). Malware detection by
ASIA. eating a whole exe. CoRR, abs/1710.09435.
Kilgallon, S., Rosa, L. D. L., and Cavazos, J. (2017). Im- Rigaki, M. and Garcia, S. (2018). Bringing a gan to a knife-
proving the effectiveness and efficiency of dynamic fight: Adapting malware communication to avoid de-
malware analysis with machine learning. In 2017 Re- tection. In 2018 IEEE Security and Privacy Work-
silience Week (RWS), pages 30–36. shops (SPW), pages 70–75.
Kirat, Jiyong, and Stoecklin (2018). Deeplocker concealing Selcuk, A. A., Orhan, F., and Batur, B. (2017). Undecidable
targeted attacks with ai locksmithing. problems in malware analysis. In 2017 12th Inter-
Kolosnjaji, B., Demontis, A., Biggio, B., Maiorca, D., Gi- national Conference for Internet Technology and Se-
acinto, G., Eckert, C., and Roli, F. (2018). Adversarial cured Transactions (ICITST), pages 494–497.
malware binaries: Evading deep learning for malware Shirataki, S. and Yamaguchi, S. (2017). A study on in-
detection in executables. CoRR, abs/1803.04173. terpretability of decision of machine learning. In
Lim, H., Yamaguchi, Y., Shimada, H., and Takakura, H. 2017 IEEE International Conference on Big Data (Big
(2015). Malware classification method based on se- Data), pages 4830–4831.
quence of traffic flow. In 2015 International Con- Soliman, S. W., Sobh, M. A., and Bahaa-Eldin, A. M.
ference on Information Systems Security and Privacy (2017). Taxonomy of malware analysis in the iot. In
(ICISSP), pages 1–8. 2017 12th International Conference on Computer En-
M. Chess, D. and R. White, S. (2000). An undetectable gineering and Systems (ICCES), pages 519–529.
computer virus. In Proceedings of Virus Bulletin Con- Su, J., Vasconcellos, V. D., Prasad, S., Daniele, S., Feng,
ference. Y., and Sakurai, K. (2018). Lightweight classifica-
Magnusardottir, A. (2018). Fileless ransomware: How it tion of iot malware based on image recognition. In
works & how to stop it? White Paper. 2018 IEEE 42nd Annual Computer Software and Ap-
Martinelli, F., Mercaldo, F., Saracino, A., and Visaggio, plications Conference (COMPSAC), volume 01, pages
C. A. (2016). I find your behavior disturbing: Static 664–669.
and dynamic app behavioral analysis for detection of Su, M.-Y. and Fung, K.-T. (2016). Detection of android
android malware. In 2016 14th Annual Conference on malware by static analysis on permissions and sensi-
Privacy, Security and Trust (PST), pages 129–136. tive functions. In 2016 Eighth International Confer-
Miller, C. and Valasek, C. (2015). Remote exploitation of ence on Ubiquitous and Future Networks (ICUFN),
an unaltered passenger vehicle. White Paper. pages 873–875.
Moubarak, J., Chamoun, M., and Filiol, E. (2018). Devel- Yang, W., Kong, D., Xie, T., and Gunter, C. A. (2017). Mal-
oping a k-ary malware using blockchain. In NOMS ware detection in adversarial settings: Exploiting fea-
2018 - 2018 IEEE/IFIP Network Operations and ture evolutions and confusions in android apps. In AC-
Management Symposium, pages 1–4. SAC.
Naeem, H., Guo, B., and Naeem, M. R. (2018). A light- Yeo, M., Koo, Y., Yoon, Y., Hwang, T., Ryu, J., Song,
weight malware static visual analysis for iot infras- J., and Park, C. (2018). Flow-based malware de-
tructure. In 2018 International Conference on Artifi- tection using convolutional neural network. In 2018
cial Intelligence and Big Data (ICAIBD), pages 240– International Conference on Information Networking
244. (ICOIN), pages 910–913.
Narayanan, A., Yang, L., Chen, L., and Jinliang, L. Zhang-Kennedy, L., Assal, H., Rocheleau, J., Mohamed, R.,
(2016). Adaptive and scalable android malware de- Baig, K., and Chiasson, S. (2018). The aftermath of a
tection through online learning. In 2016 International crypto-ransomware attack at a large academic institu-
Joint Conference on Neural Networks (IJCNN), pages tion. In Proceedings of the 27th USENIX Conference
2484–2491. on Security Symposium, SEC’18, pages 1061–1078,
Berkeley, CA, USA. USENIX Association.
OSahn, D., Kural, O. E., Akleylek, S., and Kiliç, E. (2018).
New results on permission based static analysis for an-
droid malware. In 2018 6th International Symposium
on Digital Forensic and Security (ISDFS), pages 1–4.