© SEP 2025 | IRE Journals | Volume 9 Issue 3 | ISSN: 2456-8880
IRE 1710969 ICONIC RESEARCH AND ENGINEERING JOURNALS 1840
Improving DDoS Detection in Software-Defined
Networks Through a Hybrid Machine Learning
Approach
FRANCIS ONOJAH
1
, PROF. PREMA KIRUBAKARAN
2,
DR. RIDWAN KOLAPO
3
, DR.
TEMITOPE OLUFUNMI ATOYEBI
4
, DR. R. RENUGA DEV
5
1, 2, 3, 4
Department of Information Technology, Nile University of Nigeria.
5
Associate Professor, Department of Computer Science and Applications, Faculty of Science and
Humanities, SRM Institute of Science and Technology, Chennai Ramapuram.
Abstract- (DDoS) Attacks remain a significant concern
for network security, utilizing flood-like traffic at the
volume, protocol, and application levels to exploit
vulnerabilities in today's infrastructure. To lessen these
risks, Software-Defined Networking (SDN) offers
programmability and centralized control. However,
current machine learning (ML)-based detection
techniques have a high false positive rate, are not very
flexible against zero-day attacks, and are ineffective when
handling high-dimensional flow data. To enhance the
detection of DDoS attacks in software-defined networks,
this paper proposes a hybrid machine-learning approach.
Tapping into SDNs broad view of all network flows, the
system studies traffic in real time by merging supervised
deep learning- in this case, Long Short-Term Memory-
with unsupervised anomaly detection called Isolation
Forest. The LSTM sorts incoming packets and learns new
normal behavior, while the Isolation Forest flags any
stray patterns that don’t fit.
Keywords: DDoS attacks, network security, Long Short-
Term Memory (LSTM), CNN
I. INTRODUCTION
The magnitude and sophistication of Distributed
Denial of Service (DDoS) cyber operations have
grown in magnitude along with the significant
proliferation of cloud services, Internet of Things
(IoT) devices, and real-time applications. Attackers
now employ botnets, reflection protocols, and
encrypted traffic to overwhelm their targets and
create a multitude of issues ranging from disruption
of service, economic losses, and reputational impact.
The average cost of a DDoS attack is more than $2.5
million, according to a 2023 IBM report, highlighting
the necessity of strong detection and mitigation
systems [1]. Because of their manual configurations,
rigid infrastructure, and decentralized control,
traditional network architectures find it difficult to
fend off these threats. A paradigm shift was brought
about by Software-Defined Networking (SDN),
which separated the data plane (switches) from the
control plane (centralized controller) to allow for
dynamic policy enforcement and programmable
traffic management [2]. While the worldwide
network visibility of SDN and OpenFlow-based flow
monitoring offers inherent advantages for security, its
centralized architecture also introduces new attack
surfaces. For instance, attackers can saturate control-
plane bandwidth, overflow flow tables, or spoof
source IPs to disrupt legitimate traffic. Threshold-
based techniques were used for early DDoS detection
in SDN [3]. These methods, however, are unable to
identify new or adaptive attacks, like slow and low-
speed HTTP floods or traffic patterns produced by
artificial intelligence. The ability of machine learning
(ML) to analyze high-dimensional flow data, such as
packet counts and flow durations, and spot minute
irregularities has made it popular. The application of
supervised models, including Random Forest and
Support Vector Machines (SVM), demonstrated
acceptable accuracy; however, their efficacy is
contingent upon the availability of labeled datasets,
and they remain susceptible to zero-day attacks. (85–
94%). They can and do pose issues in dynamic
environments and against zero-day attacks.
Unsupervised methods, require no labeled datasets,
like K-means clustering or autoencoders, and develop
models to learn normal traffic baseline states for an
hour for a single user account. They are uninformed
and open to misclassifications and a high false
positive rate.
To combine the advantages of supervised and
unsupervised learning, recent research investigated
hybrid machine learning models. For instance, [4]
achieved 96% accuracy in IoT intrusion detection by
combining Random Forest and K-means clustering.
These models, however, are not tailored to the