Efficient and Secure Statistical DDoS Detection Scheme
Hussein Majed
1
, Hassan N. Noura
1,2
, Ola Salman
2
, Mohammad Malli
1
and Ali Chehab
2
1
Arab Open University, Department of Computer Sciences, Beirut, Lebanon
2
American University of Beirut, Department of Electrical and Computer Engineering, Lebanon
Keywords:
DDoS, Intrusion Detection, Traffic Aggregation, Network Security.
Abstract:
One of the hardest challenges in cybersecurity is the detection and prevention of Distributed Denial of Service
(DDoS) attacks. In this paper, a lightweight statistical approach for DDoS detection is presented, in addition
to preventive and corrective countermeasures. The proposed solution is designed to be applied at the Internet
Service Provider (ISP) level. Based on aggregated NetFlow statistics, the proposed solution relies on the Z-
score and co-variance measures to detect DDoS traffic as a deviation from normal traffic. The implementation
results show a high detection rate (up to 100%) for 30 seconds time slot.
1 INTRODUCTION
Cyber threats are becoming more sophisticated and
more severe than ever before, taking advantage of ex-
isting security vulnerabilities such as the enormous
network scale, the heterogeneity of the adopted com-
munication protocols and connected devices, and the
wide deployment of applications. Organizations are
relying on different security techniques to protect
their systems and networks against cyber attacks to
minimize the associated financial losses.
Distributed Denial of Service (DDoS) attacks are
among the most disruptive threats in the cyber world.
DDoS is a variant of DoS attacks, where the attacker
attempts to make a device or network resources un-
available to its legitimate users, by temporarily or per-
manently disrupting its services. DoS attacks can be
divided into two types (Zargar et al., 2013; Gulihar
and Gupta, 2020): the first type is when an attacker
mystifies the protocol or application running on the
victim’s machine by sending malicious packets (vul-
nerability attack). The second one is when the at-
tacker sends huge traffic to take up all the resources
of the victim’s machine, making it unable to inter-
act with legitimate users (bandwidth/flooding attack).
The difference between DoS and DDoS attacks is that
the former uses only one single entity as a source,
while the latter uses a distributed set of source de-
vices. Also, DDoS attacks can be launched against
one or several victims (Douligeris and Mitrokotsa,
2004). The distributed nature of DDoS attack makes
them more devastating compared to DoS attacks.
In February 2000, a high-profile DDoS attack was
presented by targeting popular websites such as ebay,
Yahoo, Buy.com, Amazon.com, Excite.com, and Ca-
ble News Network (CNN) (Radware, 2017). Unfortu-
nately, the number of DDoS incidents is continuously
growing. Recently, DDoS attackers have been using
techniques that are very hard to detect, and the attack
size and frequency are rising rapidly (Zargar et al.,
2013). A recent major attack was recorded in 2016,
reaching approximately 1 Tbps and targeting a host-
ing cloud computing company by using more than
152,000 Internet of Things (IoT) devices (Bertino and
Islam, 2017; Kolias et al., 2017; De Donno et al.,
2018)). Because of the drastic impacts of DDoS at-
tacks, researchers are continuously trying to build
efficient, robust and comprehensive DDoS detection
and mitigation solutions (Bhatia et al., 2018). De-
fending against this type of attacks is crucial due to its
negative impact on online services availability, espe-
cially that almost every sector is becoming connected
to the Internet, including governmental, health, bank-
ing, etc.
Current DDoS detection solutions are based on
statistical or machine learning approaches. However,
these solutions suffer from performance and/or accu-
racy issues, as explained in Section 2. Hence, there is
a pressing need for a reliable and efficient solution to
respond better to these security challenges. This work
focuses on the early detection of DDoS attacks at the
ISP level. The main goal is to achieve a high level
of efficiency and speed with minimum computational
complexity. Thus, we present a fast and reliable solu-
tion based on the NetFlow statistics at edge routers of
an ISP network.
The rest of the paper is organized as follows. Sec-
tion 2 reviews the different approaches used in DDoS
Majed, H., Noura, H., Salman, O., Malli, M. and Chehab, A.
Efficient and Secure Statistical DDoS Detection Scheme.
DOI: 10.5220/0009873801530161
In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - DCNET, OPTICS, SIGMAP and WINSYS, pages 153-161
ISBN: 978-989-758-445-9
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
153
detection. Section 3 presents the proposed solution
for DDoS detection including data collection, data
aggregation and features extraction, attack detection,
and attack mitigation. Section 4 presents the exper-
imental setup and results. Finally, we conclude the
paper in Section 5.
2 RELATED WORK
The various sub-classes of DDoS detection solutions
are described in this section, mainly the ones based
on the statistical and machine learning approaches.
2.1 Statistical Approach
In the statistical approach for DDoS detection, traffic
statistics are extracted, and these include the number
of destination/source Internet Protocol (IP) addresses,
Transmission Control Protocol (TCP) flags, packet
sizes, flow rate, and others, to identify any abnormal
traffic behavior.
In (Hofstede et al., 2013), a lightweight intrusion
detection solution was presented using three metrics,
the number of flows per second, the number of bytes
per flow, and the number of packets per flow. Two
algorithms were applied on a dataset captured on the
backbone link of the Czech national research and ed-
ucation network CESNET. The first algorithm is the
Exponentially Weighted Moving Average (EWMA)
for mean calculation, extended by thresholds and cu-
mulative sum (CUSUM), and the second algorithm
is a combination of the first algorithm with season-
ality modeling. Both algorithms rely on EWMA for
calculating the mean over the past values. The re-
sults showed that a response time of 5 seconds can be
achieved with a reduction in detection delays to about
10% compared to the Netflow Sensor (NfSen) tool,
which uses time slots of 5 minutes, resulting into an
average delay of 150 seconds. In fact, the delay is
calculated based on the difference between the time a
packet is metered and the time flow data is made avail-
able to analysis. The distance-based statistical ap-
proach was introduced in (You et al., 2007) for DDoS
detection using two techniques: the average distance
estimation and the distance-based traffic separation.
The DDoS attack detection is based on analyzing dis-
tance values and traffic rates by computing the Mini-
mum Mean Square Error (MMSE) of the traffic rates
from different distances. However, one limitation of
the proposed method is that detection is unfeasible
when final TTL values of all packets are equal be-
cause the distance values, based on TTL, will be the
same for all packets. In (Girma et al., 2015), mul-
tiple statistical approaches such as co-variance ma-
trix, Kendall’s Tau, and entropy were evaluated and
a hybrid method was proposed based on entropy and
co-variance matrices. Finding the best threshold that
separates normal and abnormal traffic is the hardest
part for statistical-based DDoS detection. In (David
and Thomas, 2015), David et al. proposed an adap-
tive threshold algorithm to update the threshold based
on traffic conditions. The authors used fast entropy to
calculate the mean and standard deviation of the flow
count during a particular time interval, and the differ-
ence between the mean value and the fast entropy, to
decide if a DDoS attack is taking place.
While most of the proposed DDoS detection so-
lutions rely on the flow source IP address, attackers
can spoof these addresses, yet they cannot control
the hop counts. This key point motivated the authors
in (Shamsolmoali and Zareapoor, 2014) to propose a
model based on hop counts and the frequency count
of each packet. The model has the ability of quickly
detecting DDoS attacks with minimum storage over-
head; it showed a 97% accuracy with minimum false
alarms.
Passively monitoring abrupt changes, in network
traffic fractal parameters, was also used to detect ab-
normal changes in network traffic. In (Xia et al.,
2012), an auto-regressive system was used to esti-
mate the Fractal dimension D and Hurst parameter
H of normal traffic. The proposed method relies on
the maximum likelihood estimate-based detection ap-
proach in order to determine the change point of pa-
rameters D and H. The changes in these two points
indicate the occurrence of a DDoS flood attack.
Other methods rely on the Domain Name Servers
Block List (DNSBL) to check if the traffic originated
from sources with bad reputations. This approach
consists of studying the flow source in a network to
detect the Weird Host List (WHL) (Sawaya et al.,
2011). Analysing the destination port for each en-
tity in the WHL aims to check if the port is receiv-
ing many connections from unknown hosts. The re-
searchers used DNSBLs to check if the entities in the
WHL are blacklisted. The method showed 90% de-
tection rate, but it needs synchronization of DNSBLs.
2.2 Machine Learning Approach
Recently, machine learning techniques have been
adopted for the detection of network attacks. The
machine learning models are used to extract pat-
terns from the network traffic, and they can be used
for clustering and detecting abnormalities. In (Zekri
et al., 2017), Zekri et al. proposed a DDoS detection
system based on the C4.5 algorithm for classification,
WINSYS 2020 - 17th International Conference on Wireless Networks and Mobile Systems
154
and Naive Bayes was applied for anomaly detection,
whereas Snort was used for signature-based attack de-
tection. Four classes were considered in the proposed
system: normal, TCP SYN attack, User Datagram
Protocol (UDP) attack, and Internet Control Mes-
sage Protocol (ICMP) attack. High detection and effi-
ciency rates (up to 98%) were achieved. In (Alkasass-
beh et al., 2016), Alkasassbeh at al. compared differ-
ent machine learning methods, Multi Layer Percep-
tron (MLP), Random Forest (RF), and Na
¨
ıve Bayes
(NB), on a dataset including HTTP flood, SQL Injec-
tion DDoS (SIDDOS), UDP flood, and smurf attacks.
The obtained accuracy results were 98.63%, 98.02%
and 96.91% for MLP, RF and NB, respectively. How-
ever, RF and NB failed to show good rates for the
Smurf class, while MLP achieved a high precision
rate. In (Hou et al., 2018), Hou et al. applied RF
on NetFlow for DDoS detection. The results had 99%
true positives and 0.5% false positives. Deep learning
was also applied for DDoS detection. An approach
called DeepDefense was proposed to detect DDoS at-
tacks in (Yuan et al., 2017). The method is based on
Recurrent Neural Network (RNN) by feeding histori-
cal information to the RNN in the aim of recognizing
repeated patterns and locating them. The method was
evaluated based on a dataset recorded for 7 days and
containing DDoS attacks in 2 days out of the 7. The
experimental results demonstrated that DeepDefense
reduces the error rate by 39.69% for the first dataset
and from 7.517% to 2.103% for the second, compared
with traditional machine learning methods. For ab-
normality detection, the Modified Global K-means al-
gorithm (MGKM) was considered as an incremental
clustering algorithm to identify clusters in (Zi et al.,
2010). The linear correlation coefficient for feature
ranking was used to recalculate the clusters. The
proposed method proved to be effective and adaptive
in detecting the different phases of DDoS. Entropy-
based approaches were also considered for DDoS de-
tection. In (Singh et al., 2018), Singh et al. consid-
ered normalized router entropy, which is the overall
probability distribution of the captured flow for some
time window, the router entropy for a particular flow
(calculated by combining the entropy of distinct flows
during the time window), packet rate, and entropy
rate to measure the growth rate of the entropy of any
random process, to define the threshold of the attack
detection. For a series of n random variables, they
calculated the entropy rate (ER) of a stochastic pro-
cess (xi). The threshold values used in the algorithm
were evaluated offline by considering the malicious
and legitimate traffic flows. Another hybrid approach
using entropy, with Support Vector Machine (SVM),
was proposed in (Hu et al., 2017) to measure network
features. The SVM classifier was applied to iden-
tify network anomalies, and the experimental results
showed minimal overhead and high accuracy. Simi-
larly, in (Yun Liu et al., 2010), SVM and Traffic Fea-
ture Conditional Entropy (TFCE) were considered for
DDoS detection. The proposed solution is based on
analyzing the multiple-to-one relation between the at-
tacked IP and the source IP addresses. SVM was ap-
plied to identify DDoS attacks based on the output of
the entropy extracted by TFCE. The experimental re-
sults showed an accuracy of 93% for DDoS detection,
within a time window of 5 seconds.
Detecting DDoS attacks is a complicated task due
to the variety of attack types (TCP flooding attack,
UDP flooding attack, ICMP flooding attack, etc.).
Addressing this problem requires real traffic and up-
to-date datasets before applying any approach for de-
tection. A common limitation of the existing research
works is the quality of datasets that have been used
during the experiments. Most of these datasets are
old or based on simulating attacks in virtual labs. An-
other issue that should be addressed is the speed of
detection in the real world scenario because apply-
ing machine learning or statistical methods on current
datasets will indicate the time taken for detection after
collection, when the time taken for collecting traffic
should be added as well. In this paper, a lightweight
and high-speed statistical approach is proposed for
DDoS detection. A real and new dataset was collected
in an ISP network. NetFlow statistics are collected
and aggregated to lessen the time overhead and com-
plexity. Thus, the proposed method is easy to deploy,
scalable, robust, and flexible. It is also efficient to
detect well known and unknown (“Zero-day”) attacks
with low false alarm rate. It exhibits an adaptive tech-
nique for the definition of the corresponding rules and
thresholds values. It incurs a low computational cost
and resources compared to existing solutions.
3 PROPOSED DDoS DETECTION
SCHEME
The proposed DDoS attack detection method consists
of 4 main steps: data collection, data aggregation,
DDoS detection, and mitigation (see Figure 1).
3.1 Data Collection
In this step, the incoming network traffic at the ISP
is collected at regular intervals, such as one minute.
This can be achieved by using a software installed at
the edge router or by using an external monitoring de-
Efficient and Secure Statistical DDoS Detection Scheme
155
Figure 1: Proposed DDoS detection scheme.
Table 1: Table of notations.
Symbol Definition
µ Mean
σ Standard deviation
CV
SA
The coefficient variation of the
number of source addresses per des-
tination IP
CV
TCP
The coefficient variation of a se-
quence of TCP flows
CV
UDP
The coefficient variation of a se-
quence of UDP flows
CV
ICMP
The coefficient variation of a se-
quence of ICMP flows
ZSC
SA
The Z-score of the number of flows
per destination IP
ZSC
TCP
The Z-score of the number of TCP
flows per destination IP
ZSC
UDP
The Z-score of the number of UDP
flows per destination IP
ZSC
InputByte
The Z-score of the sum of incoming
bytes per destination IP
ZSC
TimeDuration
The Z-score of the time duration per
destination IP
ZSC
ICMP
The Z-score of the number of ICMP
flows per destination IP
NIP The number of unique destination
IP
vice. The collected NetFlow statistics will be aggre-
gated in the next step.
3.2 Data Aggregation and Features
Extraction
At this stage, the collected NetFlow statistics are ag-
gregated per destination IP. NetFlow gives informa-
tion about the flows including: source IP address, des-
tination IP address, source port number, destination
port number, duration, number of packets, number of
bytes, and IP protocol. The collected statistics at the
first step are aggregated in a custom manner. As such,
the proposed aggregation method calculates the fol-
lowing features for each destination address:
The sum of received bytes
The sum of received packets
The count of source addresses who communicated
with each destination address
The sum of time duration
The count of UDP Flows
The count of TCP Flows
The count of ICMP Flows
The count of ”S” Flag
The count of ”F” Flag
The count of ”R” Flag
The count of ”U” Flag
The count of ”A” Flag
The count of ”P” Flag
Then, the aggregated statistics are analyzed using
the Z-score metric (see Algorithm 1. The basic Z-
score formula for a sample x is:
z =
xµ
σ
(1)
where µ and σ represent the mean and the standard
deviation of the tested samples, respectively.
The Z-score of several features (number of
TCP/UDP/ICMP flows per destination IP, sum of in-
coming packets per destination IP, sum of incoming
bytes per destination IP, time duration per destination
IP, count of source IP per destination IP) are com-
puted for each address. These values are used as input
to the next step, which can be seen as a form of non-
reversible compression. The advantage of this step
is to reduce the required processing complexity and
storage overhead.
3.3 DDoS Detection
For each aggregated flow, the obtained Z-scores of
three features are compared to the corresponding
threshold values that are based on the coefficient of
variation (CV). CV represents a statistical indicator
for the dispersion of data points in a data set (series)
WINSYS 2020 - 17th International Conference on Wireless Networks and Mobile Systems
156
around the mean. It is the ratio of the standard devi-
ation to the mean, and it is a useful statistical metric
for comparing the degree of variation from one data
series to another, even if the means are different. CV
is computed as illustrated in the following equation:
CV =
σ
µ
(2)
Accordingly, a DDoS attack is detected if the ob-
tained Z-score values of these features are greater or
equal to the corresponding thresholds (see lines 14, 17
or 20 of Algorithm 1). These thresholds are set based
on a training step (initial step), which depends on the
ISP profile (traffic and network characteristics). Us-
ing statistical thresholds is not practical because they
need to be updated manually, each time a bandwidth
upgrade takes place. In addition, they might lead to
false positives and false negatives. To make the pro-
posed solution based on a dynamic threshold, the Z-
score value of each feature is compared to its Coeffi-
cient of Variation (CV). In case of an attack, the at-
tacked IP address will definitely receive a large num-
ber of packets from multiple sources within a time
window. The attack will increase the mean value
and standard deviation for each affected feature of the
whole set, resulting into a high Coefficient of Vari-
ation for each affected feature. By calculating the
Z-score of each feature for each data point, the data
point of the attacked IP will have a high z-score value
for each affected feature, indicating that it is relatively
different. Since we have the CV value of each feature,
we can compare it to the z-score of each data point;
if the z-score is higher than the CV, it’s considered an
outlier compared to the rest of the flows.
Receiving a high number of packets or bandwidth
from a high number of source IPs on a specific IP will
definitely make the standard deviation and mean val-
ues high for each addressed feature. The Z-score is a
well known method for the detection of outliers. Ap-
plying the Z-score on each value in the aggregated
flow will clearly reveal the outlier, but each type of
an attack has its own characteristics. TCP flood at-
tacks are fast, from multiple sources, and share the
same flag. UDP flood attacks rely on high incoming
bytes, from multiple sources, and all the packets are
destined to same IP. ICMP flood attacks try to over-
whelm the targeted device ability to respond to the
high number of requests and/or overload the network
connection with bogus traffic, it is fast, and the traffic
will be destined to the same IP from probably mul-
tiple sources. Thus, the following detection features
are selected and used for each DDoS attack type:
TCP: time duration, number of TCP packets, and
number of Source IP.
UDP: sum of incoming bits, number of UDP
packets, and number of source IP.
ICMP: sum of incoming bits, number of ICMP
packets, and number of source IP.
Figure 2: Mitigation process.
3.4 Mitigation Process
When a DDoS attack is detected, the prevention coun-
termeasure is activated to mitigate it. The proposed
mitigation process is illustrated in Figure 2, and it can
be realized locally by blocking the incoming traffic
for a certain period of time (depending on the DDoS
attack scenario such as TCP or UDP), or by using the
scrubbing server (high volume traffic case). Indeed,
if the attack traffic ( 80%) of the total traffic and
the router CPU 40% of its total capacity, the attack
cannot be mitigated by blocking the incoming traffic,
but by advertising the attacked sub-net on the Border
Gateway Protocol (BGP) to the scrubbing server.
4 EXPERIMENTAL SETUP AND
RESULTS
In this section, we explain the implementation of the
proposed DDoS detection and prevention scheme in
detail.
4.1 NetFlow Collection
For collecting data, the setup was realized in an ISP
network with full traffic visibility. NetFlow is a fea-
ture introduced by Cisco to enable in/out IP network
traffic statistics collection at the router interface(s).
Efficient and Secure Statistical DDoS Detection Scheme
157
Algorithm 1: Proposed detection algorithm.
Input: Aggregated Flow AF
Output: R
1: procedure R = Detection(AF)
2: NIP Numbero f Rows(AF)
3: CV
SA
std(C
SA
)
mean(C
SA
)
4: CV
TCP
std(C
TCP
)
mean(C
TCP
)
5: CV
UDP
std(C
UDP
)
mean(C
UDP
)
6: CV
ICMP
std(C
ICMP
)
mean(C
ICMP
)
7: for i 0 to NIP do
8: ZSC
SA
Zscore(C
SA
)
9: ZSC
TCP
Zscore(C
TCP
)
10: ZSC
UDP
Zscore(C
UDP
)
11: ZSC
ICMP
Zscore(C
ICMP
)
12: ZSC
InputByte
Zscore(C
InputByte
)
13: ZSC
TimeDuration
Zscore(C
TimeDuration
)
14: if |ZSC
SA
| CV
SA
& &|ZSC
TCP
| CV
TCP
& &ZSC
TimeDuration
0 then
15: TCP DDoS Attack Detected on DA
16: end if
17: if |ZSC
SA
| CV
SA
& &|ZSC
UDP
| CV
UDP
& &|ZSC
InputByte
| CV
InputByte
then
18: UDP DDoS Attack Detected on DA
19: end if
20: if |ZSC
SA
| CV
SA
& &|ZSC
ICMP
| CV
ICMP
& &|ZSC
InputByte
| CV
InputByte
then
21: ICMP DDoS Attack Detected on DA
22: end if
23: end for
24: end procedure
A virtual CentOS machine was created with vmware
ESXi 6.0. This machine has 2 network interfaces, the
first one is for NetFlow collection, and the second one
is for traffic management. To isolate the communi-
cation between this machine and the edge routers, a
separate VLAN was created. Two Cisco edge routers
(ASR 9K) running IOS XR Release 7.0.1 were con-
figured to send all incoming NetFlow traffic to the
CentOS machine.
”nfcapd” was installed on the CentOS machine to
capture the NetFlow information. Multiple instances
of ”nfcapd” were configured to listen on different
ports and to write the captured flows, each minute,
in a separate file, and for each router, in a separate
folder.
nfcapd -w -D -l /flow base dir/router1 -p 23456
nfcapd -w -D -l /flow base dir/router1 -p 23456
NetFlow statistics of both routers were collected
for one month. During this period, multiple DDoS at-
tacks hit the ISP network. A time table was updated
each time an attack occurred to be used later on to dif-
ferentiate the benign traffic from the attack one. The
attack usually saturates the international links of the
ISP and increases the CPU usage on edge routers.
The ”nfcapd” tool saves the files in raw format,
and the ”nfdump” tool is applied to read the input file,
and to provide statistics, aggregation, searching, sort-
ing, and conversion to another format such as CSV.
Thus, the conversion option was used to convert the
raw files to CSV files, which represent the dataset
used for evaluating the proposed detection method.
4.2 Data Aggregation
The captured flows can be visualized using ”nfdump”
statistics feature by aggregating the traffic by destina-
tion IP and sorting the results by number of packets
received for each destination IP. The results showed
that the attacked IP always had the highest number of
packets during the specified time duration (here one
minute) (See Figure 3).
Examining the traffic destined to the IP with
the highest number of received packets, the results
showed that, for UDP traffic, this IP receives traffic
from more than 40,000 different IPs on a specific port,
during one minute. Repeating the same test on multi-
ple files, containing different types of attacks, the out-
WINSYS 2020 - 17th International Conference on Wireless Networks and Mobile Systems
158
Figure 3: NetFlow aggregation using nfdump.
Figure 4: False positive VS time.
come was similar, and the only difference was in TCP
attacks where the flows shared the same TCP flag
(SYN flooding, Ack flooding, etc.). Figure 3 shows
that the statistics provided by ”nfdump” are limited
and not enough in order to detect the abnormality in
the NetFlow traffic, because the destination port and
many fields that might be useful for DDoS detection
are missing. Although NetFlow aggregation using
”nfdump” presents promising results in detecting out-
liers, yet some features such as the count of source
IP addresses, the sum of UDP/ICMP/TCP packets are
key for DDoS detection. Consequently, a customized
aggregation method of NetFlow is needed for detect-
ing various types of DDoS attacks. Another chal-
lenge is the use of the appropriate statistical approach
on the customized aggregated flow to detect the out-
liers. This was considered by our proposed NetFlow
aggregation method, which extracts main features for
DDoS detection. The Z-score and co-variance are cal-
culated for the extracted features to detect DDoS as
shown in the next section.
4.3 DDoS Detection
Identifying the best time window to detect attacks is
very difficult. For this purpose, the proposed solu-
tion was tested for different time slots. The collected
datasets were aggregated based on 10, 30, and 60 sec-
onds time windows. The results, shown in Figure 5
for TCP DDoS attacks and in Figure 6 for UDP DDoS
attacks, clearly indicate that using the coefficient of
variation as a dynamic threshold is efficient for cor-
rectly detecting DDoS attacks. For TCP attacks, the
method worked perfectly for 10, 30, and 60 seconds
time windows. For UDP attacks, as shown in Fig-
ure 6, some destination IPs had the number of source
addresses greater than the threshold (blue dots), but
other features (number of bytes and number of pack-
ets ) are not higher than the thresholds. Thus, these
flows are not detected as attacks. To lower the de-
tection time, we studied the minimum required delay
(seconds) to detect DDoS attacks. Figure 4 presents
the variation of false-positives versus the time for
several DDoS cases (each color represents a type of
DDoS attack). According to the results in Figure 4, a
time window greater or equal to 30 seconds showed
zero false positives and 100% true positives for TCP
DDoS attacks. Note that for a time window of 10
seconds, several of these attacks were detected, but
other ones were not detected. Therefore, the results
indicated that the proposed method is efficient if ap-
plied on a time window of 30 seconds, but some cru-
cial real-time applications and high sensitive servers
might get harmed if the attack lasts for 30 seconds
Efficient and Secure Statistical DDoS Detection Scheme
159
(a) DDoS detection within 10 seconds (b) DDoS detection within 30 seconds (c) DDoS detection within 60 seconds
Figure 5: TCP DDoS detection within 10, 30, and 60 seconds.
(a) DDoS detection within 10 seconds (b) DDoS detection within 30 seconds (c) DDoS detection within 60 seconds
Figure 6: UDP DDoS detection within 10, 30, and 60 seconds.
before being detected. For this reason, the maximum
Z-score of each feature, in the datasets showing false
positive alerts, was studied. We concluded that the
threshold of Z-score, which is the coefficient of vari-
ation, should not be less than 17 for sensitive services
that need detection within 10 seconds. The script
should ignore the alert if the CV value is less than
17 since these are most likely due to false positives.
4.4 DDoS Mitigation
The proper mitigation technique should be based on
the type and the volume of the attack. For this pur-
pose, a workflow for mitigation was created in Fig-
ure 2. If the attack volume and severity are accept-
able, the mitigation can be done locally by blocking
the attack traffic on the edge router of the ISP before
reaching the customers network. However, if the at-
tack severity and volume are huge, the proposed so-
lution forces the edge router to advertise the attacked
sub-net on BGP to the scrubbing server, which blocks
the attack traffic and sends the benign traffic back to
the ISP using Generic Routing Encapsulation (GRE)
tunnels. Cleaning the traffic locally, if possible, is bet-
ter than sending it to scrubbing servers since the GRE
tunnel increases the delay and consumes 5% of the
traffic.
5 CONCLUSION
DDoS attacks are very serious security threats that
need to be detected and prevented in current and fu-
ture networks. In this paper, we proposed detective,
preventative and corrective measures against inbound
DDoS attacks. The proposed solution was designed
at the ISP level, and it is based on a lightweight sta-
tistical approach. Technically, the proposed method
applies first aggregation on NetFlow, then statistical
analysis to detect DDoS traffic as outliers by using
the Z-score and co-variance variation metrics. The
proposed technique requires 30 seconds to be confi-
dent that a DDoS attack is taking place. The main
target of this work is to solve the issue of DDoS at-
tacks by reducing the required resources and latency,
which is considered as one of the hardest obstacles to
overcome in cybersecurity.
REFERENCES
Alkasassbeh, M., Al-Naymat, G., Hassanat, A., and Alm-
seidin, M. (2016). Detecting distributed denial of ser-
vice attacks using data mining techniques. Interna-
tional Journal of Advanced Computer Science and Ap-
plications, 7(1):436–445.
Bertino, E. and Islam, N. (2017). Botnets and internet of
things security. Computer, 50(2):76–79.
Bhatia, S., Behal, S., and Ahmed, I. (2018). Distributed de-
nial of service attacks and defense mechanisms: Cur-
WINSYS 2020 - 17th International Conference on Wireless Networks and Mobile Systems
160
rent landscape and future directions. In Versatile Cy-
bersecurity, pages 55–97. Springer.
David, J. and Thomas, C. (2015). Ddos attack detection us-
ing fast entropy approach on flow-based network traf-
fic. Procedia Computer Science, 50:30–36.
De Donno, M., Dragoni, N., Giaretta, A., and Spognardi,
A. (2018). Ddos-capable iot malwares: Comparative
analysis and mirai investigation. Security and Com-
munication Networks, 2018.
Douligeris, C. and Mitrokotsa, A. (2004). Ddos attacks and
defense mechanisms: classification and state-of-the-
art. Computer Networks, 44(5):643–666.
Girma, A., Moses, G., Li, J., Lui, C., and Abayomi, K.
(2015). Analysis of ddos attacks and an introduction
of a hybrid statistical model to detect ddos attacks on
cloud computing environment.
Gulihar, P. and Gupta, B. (2020). Cooperative mechanisms
for defending distributed denial of service (ddos) at-
tacks. In Handbook of Computer Networks and Cyber
Security, pages 421–443. Springer.
Hofstede, R., Barto
ˇ
s, V., Sperotto, A., and Pras, A. (2013).
Towards real-time intrusion detection for netflow and
ipfix. In Proceedings of the 9th International Con-
ference on Network and Service Management (CNSM
2013), pages 227–234. IEEE.
Hou, J., Fu, P., Cao, Z., and Xu, A. (2018). Machine learn-
ing based ddos detection through netflow analysis. In
MILCOM 2018-2018 IEEE Military Communications
Conference (MILCOM), pages 1–6. IEEE.
Hu, D., Hong, P., and Chen, Y. (2017). Fadm:
Ddos flooding attack detection and mitigation sys-
tem in software-defined networking. In GLOBE-
COM 2017-2017 IEEE Global Communications Con-
ference, pages 1–7. IEEE.
Kolias, C., Kambourakis, G., Stavrou, A., and Voas, J.
(2017). Ddos in the iot: Mirai and other botnets. Com-
puter, 50(7):80–84.
Radware (2017). Ddos attack history radware security.
https://security.radware.com/ddos-knowledge-center/
ddos-chronicles/ddos-attacks-history/. (Accessed on
05/17/2020).
Sawaya, Y., Kubota, A., and Miyake, Y. (2011). Detection
of attackers in services using anomalous host behav-
ior based on traffic flow statistics. In 2011 IEEE/IPSJ
International Symposium on Applications and the In-
ternet, pages 353–359.
Shamsolmoali, P. and Zareapoor, M. (2014). Statistical-
based filtering system against ddos attacks in cloud
computing. In 2014 International Conference on Ad-
vances in Computing, Communications and Informat-
ics (ICACCI), pages 1234–1239.
Singh, K., Dhindsa, K. S., and Bhushan, B. (2018).
Threshold-based distributed ddos attack detection in
isp networks. Turkish Journal of Electrical Engineer-
ing & Computer Sciences, 26(4):1796–1811.
Xia, Z., Lu, S., and Li, J. (2012). Ddos flood attack de-
tection based on fractal parameters. In 2012 8th In-
ternational Conference on Wireless Communications,
Networking and Mobile Computing, pages 1–5.
You, Y., Zulkernine, M., and Haque, A. (2007). Detecting
flooding-based ddos attacks. In 2007 IEEE Interna-
tional Conference on Communications, pages 1229–
1234. IEEE.
Yuan, X., Li, C., and Li, X. (2017). Deepdefense: identify-
ing ddos attack via deep learning. In 2017 IEEE Inter-
national Conference on Smart Computing (SMART-
COMP), pages 1–8. IEEE.
Yun Liu, Jianping Yin, Jieren Cheng, and Boyun Zhang
(2010). Detecting ddos attacks using conditional en-
tropy. In 2010 International Conference on Computer
Application and System Modeling (ICCASM 2010),
volume 13, pages V13–278–V13–282.
Zargar, S. T., Joshi, J., and Tipper, D. (2013). A survey
of defense mechanisms against distributed denial of
service (ddos) flooding attacks. IEEE communications
surveys & tutorials, 15(4):2046–2069.
Zekri, M., El Kafhali, S., Aboutabit, N., and Saadi, Y.
(2017). Ddos attack detection using machine learn-
ing techniques in cloud computing environments. In
2017 3rd International Conference of Cloud Comput-
ing Technologies and Applications (CloudTech), pages
1–7. IEEE.
Zi, L., Yearwood, J., and Wu, X. (2010). Adaptive clus-
tering with feature ranking for ddos attacks detection.
In 2010 Fourth International Conference on Network
and System Security, pages 281–286.
Efficient and Secure Statistical DDoS Detection Scheme
161