Deep Learning Framework for Cyber Attack Prediction: Review

This literature review examines various approaches to cyber attack prediction, focusing on the evolution from statistical models to deep learning frameworks. The review begins with an introduction highlighting the increasing prevalence and impact of cyberattacks, emphasizing the need for early prediction to mitigate risks. It then analyzes several key research papers, including studies utilizing vector autoregression (VAR) models, marked point processes, Kalman filters, and machine learning techniques like the Generalized Regression Neural Network (GEFTS–GRNN) ensemble. The analysis delves into the strengths and weaknesses of each approach, discussing their methodologies, findings, and limitations. The review highlights the shift towards deep learning, particularly the BRNN-LSTM framework, which offers improved accuracy and efficiency. The conclusion synthesizes the findings, suggesting a hybrid approach combining the strengths of statistical and deep learning methods to create a comprehensive and effective cyber attack prediction framework. The student's future research will build upon the BRNN-LSTM approach, extending it to predict various types of cyberattacks and incorporating the advantages of other techniques.

Deep Learning Framework for
Cyber Attack Prediction
Literature Review (E1)
Research Methods
Date:
Supervisor:
Name:
Student ID:

Table of Contents
1. Introduction .............................................................................................. 1
2. Literature Analysis …………………………………………………………… 2
3. Conclusions ….……………………………………………………………….. 7
References ……….……………………………………………………………… 9

1
1. Introduction
“Cyberattack” is a hot topic in today’s digitalized world. Cyber technologies have opened
new doors to organizations and people to digitize them with the latest technologies and
ease their work. Unfortunately, this has opened opportunities for illegitimate users like
hackers as well. Cyber attacks are increasing immensely and make a great loss to
organizations. For example, cyber-attacks increased by 37% over the past month with
COVID19 pandemic (Muncaster, 2020). According to Bloomberg, a cyber attack hit the
U.S. Health Agency this month to steal national health data and to publish fake
news(SteinJacobs, 2020). In January 2020, the Puerto Rico government lost $2.6
Million in a phishing scam(PRESS, 2020). These cyber-threats make monetary losses,
loss of trust, harm to people and it directly affects the reputation of the companies.
There are a lot of cybersecurity tools out there to secure cyberspace. It is always a
necessity to predict the attack in the early stages before it harms the organization. So
that the companies can set up precautions to stop the attack. For example, Honeypots
and network telescopes monitor unsolicited internet traffic to gather data related to
cyberattacks in networks and to allocate defence tools to secure the network(Peng et
al., 2016).
The subject of predicting cyber attacks and efficient models have reached a top
research topic in the last few years. Different authors followed different approaches to
predict cyberattacks while having advantages and disadvantages in their models in
different situations. The current research path moving towards top-notch technologies
like artificial intelligence, machine learning, deep learning and big data.

2. Literature Analysis
This section introduces the related work conducted in the past by different researchers.
This section includes different approaches they used, their conclusions, findings and
drawbacks.
The paper written by Ling, introduces a regression-based analytical model to predict
cyber attacks in Honeynets. Honeynets are composed of multiple honeypots. The main
purpose of conducting this study was to identify geospatial and temporal patterns in the
cyber attacks and use the knowledge for future attack predictions. The authors
introduced a vector autoregression(VAR) model for honeynets. This study used a
dataset with 9 AWS virtual honeypot hosts. The paper well explained the ways to use
VAR and BigVAR models in predictions. The researchers introduced a fractional
integration methodology to calculate Large Range Memory- LRM in each host which
helps to achieve concise modelling and high performance. Moreover, they found that
the dependency among hosts does not improve prediction accuracy if honeypot hosts
are not directly connected in the same network(Ling et al., 2019).
The research conducted by Peng(2019) about cyberattack rates used another approach
to accurately predict cyber attacks. These authors have studied extreme value
phenomenon exhibits in cyberattack rates. In short, extreme value phenomenon is the
number of attacks against a system of interest per time unit. This value is very important
when allocating defence resources by the defender to protect networks at the right time.
The authors introduced a marked point process technique to model and predict these
extreme cyber-attack rates. They have used Value-at-Risk(VaR) to measure the
intensity of the attacks over a period of time by providing the probability of extreme
cyber-attack rates with a certain confidence level. In addition to that, they have used
the Point Over Threshold(POT) method to model the magnitude of extreme attack rates.

3
Autoregressive Conditional Duration(ACD) approach is used to describe the arrival of
extreme attack rates. Final analysis results revealed that using ACD and Log_ACD give
higher precision on prediction than the FARIMA and GANCH models(Peng et al., 2016).
ARIMA, FARIMA and GANCH models are the most common conventional statistical
models used in cyber attack predictions using time series data. Autoregressive
Integrated Moving Average - ARIMA model is used to analyse time-series
data(CryerChan, n.d.). Generalized Autoregressive Conditional Heteroskedasticity -
GANCH model used for more accurate predictions by accommodating extreme value
data in a time series(CryerChan, n.d.). Fractionally Autoregressive Integrated Moving
Average - FARIMA model used to predict the cyberattacks with long-range dependent
data in a time series and this is an extended version of ARIMA(Fang et al., 2019).
When compared to the research conducted by Ling, this marked point process is used
in predicting extreme attack rates in clusters and they can accommodate the
dependence between the inter-exceedance. This model is highly capable with large and
small datasets. However, the authors assumed that 1h time gap(optimal unit of time) for
the prediction is enough for the defender to allocate defence tools to secure the
network. But this paper does not explain the ways to determine the optimal unit of time
for the predictions. It is the main future work to be done in this process. Moreover, in
this approach, the data in the dataset need to be preprocessed before using it for
predictions. For example honeypot/telescope dataset can include data created as
results of misconfigurations of the servers(Peng et al., 2016).
Another approach of predicting cyber attack is studied by Daria in 2019. He conducted
this research on industrial systems and used Kalman filters. Kalman filters are time
series analytical models which can efficiently process, recursively filter and analyse
datasets. This study was conducted on a Gasoil data set which was collected as sensor
data in the Gasoil Heating Loop process. The authors have selected industrial systems

4
for the study because detecting and predicting cyber attacks in industrial systems are
complicated due to high heterogeneity and low intelligent devices. The researchers
calculated the prediction rates using mean square error and mean absolute deviation
values and received a great result when compared with other models in industrial
systems(Daria et al., 2019).
When compared with other models described above, the usage of Kalman filter has
given a high strength to this approach. These filters can be used on both linear and
non-linear processes and use filtering recursively. This approach can reduce memory
storage problems as this approach does not require to know the history and this uses
system states for system state forecasting. This feature allows the model to adapt new
conditions effectively. In addition to that extended version of this approach can analyse
multivariate time series and forecast cyber-attacks which most of the other models are
unable to accomplish(Daria et al., 2019).
The study completed by Ghiyas Smith(2011) introduced a novel approach to predict
time series data using machine learning concepts. The reviewed papers above were
based on statistical approaches and this paper used new technologies to predict future
events. This paper is not directly written to the cybersecurity field, but it is a very
worthwhile approach to use in the cyberattack prediction.
This research presents a novel homogeneous neural network approach known as
Generalized Regression Neural Network (GEFTS–GRNN) Ensemble to forecast the
future based on time-series data. This model uses a dynamic nonlinear weighting
system which consists of base GRNNs and a combiner GRNN. General regression
neural network- GRNN is another neural network algorithm. The base-level GRNN
receives different sets of datasets with different seasonal time series patterns as inputs
and passes the output to another combiner GRNN. When GRNN itself works with
multiple predictors it gives worst performances(GheyasSmith, 2009) but this issue is

5
fixed with this GEFTS–GRNN. This approach increases the accuracy of the prediction.
Not only GRNN but also this model is a combination of several well-respected
algorithms(GheyasSmith, 2011).
The comparison between other models showed that GEFTS–GRNN is more powerful
and accurate. This model eliminated most of the drawbacks other models have like local
optima, overfitting, dimension disasters which make the algorithms inefficient. The other
models are inefficient and give low accuracy results with seasonal time series as most
of them are based on global approximators. But the proposed GEFTS approach uses
local approximators and works well with seasonal time series data. This model has
increased its prediction accuracy by having multiple neural networks(GheyasSmith,
2011).
The major disadvantage of GEFTS is, the algorithms used in this approach are complex
and need high computational power. Because of the complexities, the time to calculate
the prediction is higher than other models. Furthermore, this approach needs to
preprocess data before passing them to the input layers(GheyasSmith, 2011).
With the development of machine learning the researchers tend to use other new
technologies like deep learning to these prediction approaches. One of the main studies
conducted using deep neural networks is the introduction of BRNN-LSTM by Fang and
his team(Fang et al., 2019). This study used a novel bi-directional recurrentneural
network with long short term memory framework(BRNN-LSTM). The bi-directional
recurrent neural network(RNN) is a feed-forward network which helps the network to
train itself and increase the prediction accuracy with the frequency of usage. LSTM is
in-memory states which are used to store memory status at different nodes and they
help to increase the performance. The training process of RNN can cause gradient
vanishing problems and LSTM are used to fix it. The whole framework uses statistical
properties of cyberattack rates time-series data(Fang et al., 2019).

6
The accuracy measurement like present mean absolute deviation and mean absolute
percentage error values achieved remarkably high prediction accuracy rate for BRNN-
LSTM than other models. The data preprocessing step is avoided in the BRNN-LSTM
framework and the selection of fitted values is calculated using an algorithm.
Furthermore, the researchers conducted a comparison against other analytical
approaches with this deep learning approach and found that the deep learning
approach is more accurate and reduces error rates thanother models(NAMINAMIN,
2018).
However, the authors found that this BRNN-LSMT framework had missed the observed
values on some occasions and these occasions are not predictable. The authors
assumed the prediction accuracy as sufficient throughout the paper but this can vary
from different situations. So there is more to improve in this concept to maximize
accuracy and performance.

7
3. Conclusion
According to the literature review conducted, the cybersecurity attack prediction can be
done as illustrated in the below figure.
Figure 1: Cyber Attack prediction methodologies
According to the papers reviewed, there is a reasonable research gap to fill. The
research is almost done for conventional statistical models and the new trend is moving
towards deep learning technologies. The two papers reviewed related to machine
learning and deep learning solved most of the problems the other models had with high
prediction accuracy and performance. The framework introduced by Ghiyas Smith in
2011 found high computational costs and high complexities of the algorithms inits
neural network. Moreover, the data needs to be preprocessed before passing it to the
network. In 2019, Fang fixed all the research gaps GEFTS–GRNN had with Fang’s
BRNN-LSTM approach. In this approach, he used feedforward neural networks. Hence
the preprocess stage is avoided. Furthermore, he introduced simple but highly accurate
algorithms to predict the results. It minimizes the computational costs.

8
Fang’s research(BRNN-LSTM) opens up opportunities for my research related to Deep
Learning Framework for cyber attack prediction. BRNN-LSTM approach is mainly
focused on predicting cyberattack rates. But this research paper does notconsider
predicting cyber attacks as a whole(denial of service attack, phishing attacks etc).
BRNN-LSTM does not only rely on time series data. So Fang’s approach will be the
basement of my research topic. Even Though the neural network is the best solution for
high accuracy and higher performance solutions, there are other advantages of using
statistical approaches like Kalman filters and regression techniques in predictions. They
have specific strengths the network can use with deep neural network technologies.
These features can be implemented inside hidden layers of the network. Due to that the
network can extract more information and can identify potential attacks efficiently. Thus
the approach for Deep Learning Framework for cyber attack prediction will be a hybrid
solution based on all the strength discussed above with minimised weaknesses.

9
References
Cryer, J. and Chan, K. (n.d.) Time series analysis. 2nd ed New York: Springer, p.92-102.
Daria, L., Dmitry, Z. and Anastasiia, Y. (2019) Predicting cyber attacks on industrial systems using the
Kalman filter. 2019 Third World Conference on Smart Trends in Systems Security and Sustainablity
(WorldS4). IEEE [Online]. Available at: doi:10.1109/worlds4.2019.8904038 [Accessed: 24 March 2020].
Fang, X., Xu, M., Xu, S. and Zhao, P. (2019) A deep learning framework for predicting cyber attacks
rates. EURASIP Journal on Information Security, 2019 (1). Springer Science and Business Media LLC
[Online]. Available at: doi:10.1186/s13635-019-0090-6 [Accessed: 15 March 2020].
Gheyas, I. and Smith, L. (2009) A Neural Network Approach to Time Series Forecasting. Proceedings of
the World Congress on Engineering, 2. [Online]. Available at:
http://iaeng.org/publication/WCE2009/WCE2009_pp1292-1296.pdf [Accessed: 2 April 2020].
Gheyas, I. and Smith, L. (2011) A novel neural network ensemble architecture for time series forecasting.
Neurocomputing, 74 (18), p.3855-3864. Elsevier BV [Online]. Available at:
doi:10.1016/j.neucom.2011.08.005 [Accessed: 1 April 2020].
Ling, X., Rho, Y. and Ten, C. (2019) Predicting Global Trend of Cybersecurity on Continental Honeynets
Using Vector Autoregression. 2019 IEEE PES Innovative Smart Grid Technologies Europe
(ISGT-Europe). IEEE [Online]. Available at: doi:10.1109/isgteurope.2019.8905639 [Accessed: 4 April
2020].
Muncaster, P. (2020) Cyber-Attacks Up 37% Over Past Month as #COVID19 Bites. Infosecurity
Magazine. [Online]. Available at:
https://www.infosecurity-magazine.com/news/cyberattacks-up-37-over-past-month/ [Accessed: 2 April
2020].
NAMIN, S. and NAMIN, A. (2018) FORECASTING ECONOMIC AND FINANCIAL TIME SERIES: ARIMA
VS. LSTM. [Online]. Available at: doi:https://arxiv.org/pdf/1803.06386.pdf [Accessed: 21 March 2020].
Peng, C., Xu, M., Xu, S. and Hu, T. (2016) Modeling and predicting extreme cyber attack rates via
marked point processes. Journal of Applied Statistics, 44 (14), p.2534-2563. Informa UK Limited [Online].
Available at: doi:10.1080/02664763.2016.1257590 [Accessed: 28 March 2020].
PRESS, A. (2020) Puerto Rico Loses $2.6 Million in Phishing Scam. Courthousenews.com. [Online].
Available at: https://www.courthousenews.com/puerto-rico-loses-2-6-million-in-phishing-scam/ [Accessed:
4 April 2020].

10
Stein, S. and Jacobs, J. (2020) Bloomberg - Are you a robot?. Bloomberg.com. [Online]. Available at:
https://www.bloomberg.com/news/articles/2020-03-16/u-s-health-agency-suffers-cyber-attack-during-covi
d-19-response [Accessed: 3 April 2020].
Zhan, Z., Xu, M. and Xu, S. (2015) Predicting Cyber Attack Rates With Extreme Values. IEEE
Transactions on Information Forensics and Security, 10 (8), p.1666-1677. Institute of Electrical and
Electronics Engineers (IEEE) [Online]. Available at: doi:10.1109/tifs.2015.2422261 [Accessed: 19 March
2020].