Sensors
Sensors
Article
Enhanced Network Intrusion Detection System for Internet
of Things Security Using Multimodal Big Data Representation
with Transfer Learning and Game Theory
Farhan Ullah 1,2 , Ali Turab 1 , Shamsher Ullah 3 , Diletta Cacciagrano 2 and Yue Zhao 4, *
Abstract: Internet of Things (IoT) applications and resources are highly vulnerable to flood attacks,
including Distributed Denial of Service (DDoS) attacks. These attacks overwhelm the targeted device
with numerous network packets, making its resources inaccessible to authorized users. Such attacks
may comprise attack references, attack types, sub-categories, host information, malicious scripts,
etc. These details assist security professionals in identifying weaknesses, tailoring defense measures,
and responding rapidly to possible threats, thereby improving the overall security posture of IoT
devices. Developing an intelligent Intrusion Detection System (IDS) is highly complex due to its
numerous network features. This study presents an improved IDS for IoT security that employs
multimodal big data representation and transfer learning. First, the Packet Capture (PCAP) files
are crawled to retrieve the necessary attacks and bytes. Second, Spark-based big data optimization
algorithms handle huge volumes of data. Second, a transfer learning approach such as word2vec
Citation: Ullah, F.; Turab, A.; Ullah, S.; retrieves semantically-based observed features. Third, an algorithm is developed to convert network
Cacciagrano, D.; Zhao, Y. Enhanced bytes into images, and texture features are extracted by configuring an attention-based Residual
Network Intrusion Detection System Network (ResNet). Finally, the trained text and texture features are combined and used as multimodal
for Internet of Things Security Using features to classify various attacks. The proposed method is thoroughly evaluated on three widely
Multimodal Big Data Representation used IoT-based datasets: CIC-IoT 2022, CIC-IoT 2023, and Edge-IIoT. The proposed method achieves
with Transfer Learning and Game excellent classification performance, with an accuracy of 98.2%. In addition, we present a game
Theory. Sensors 2024, 24, 4152.
theory-based process to validate the proposed approach formally.
https://doi.org/10.3390/s24134152
Academic Editor: Alessandra Keywords: intrusion detection system; big data; transfer learning; game theory; network
Rizzardi traffic; cybersecurity
and thwart network intrusions to resolve the rising issues regarding privacy and security.
Techniques for Network Intrusion Detection Systems (NIDS) based on machine learning or
deep learning have received much attention in the scientific community [4,5]. Their aim
is to mitigate misuse or deviations within IoT infrastructures and systems. While NIDS
has proven effective in detecting malicious network activity, a significant limitation is the
inability to identify new forms of network intrusions due to the challenges posed by big data
processing and related imbalances [6]. Furthermore, existing machine learning algorithms
for intrusion detection struggle with multimodal data, emphasizing the significance of
further study into hybrid feature engineering. Big data is generated via network traffic,
system events, and logs. Cybersecurity is improved by using big data analytics and related
approaches to analyze data flows, investigate events, and regularly detect anomalies, shifts,
and threats in network traffic, as shown in Figure 1. Remote routers properly collect and
track the network traffic that IoT devices generate. These routers send such real-time data to
Spark stream processors for efficient processing. The AMP Lab at UC Berkeley developed
Spark, a platform for distributed computation [7]. A Spark-based platform can dynamically
and efficiently manage massive amounts of generated data. Due to the expansion of large
data structures, networking system speeds, and intruder techniques, intelligent intrusion
detection remains an essential research challenge. Spark incorporates several integration
techniques across multiple big data systems, enabling it to quickly handle enormous
amounts of data [8,9]. Its memory-driven Resilient Distributed Dataset (RDD) method is
designed for ensembles and recurrent computations and permits intermediary data caching
in storage. As a result, the Spark model provides particular benefits for processing massive
volumes of data through various optimization approaches, including partitioning, caching,
API selection, and more.
Figure 1. Monitoring of IoT-based network traffic flows using Spark stream processor.
Traditional Machine Learning (ML) and Deep Learning (DL) methods in NIDS have
prioritized model performance over in-depth network semantic analysis. NIDS evaluates
data from hosts and network traffic to detect intrusion attempts. Intruders regularly utilize
malicious scripts to disrupt IoT network functionality. Semantic-based feature engineering
can be applied to behavioral segmentation to detect potential malicious scripts. This
strategy entails evaluating network data semantically to generate training features before
using them for classification. Nonetheless, text-based behaviors such as code obfuscation,
insertion, and re-ordering present difficulties. In contrast, image-based feature analysis is
frequently used, capturing a wide range of structural information such as memory, process,
and header details. This enables visual images to retrieve dynamic or obfuscated data
of any type [10–12]. However, this strategy may disrupt the overall structure of network
traffic, making it harder to target specific scripts such as DDoS or authentication protocols.
It is also entirely dependent on the properties of images, which leaves it open to attacks
that impair image quality. To address these concerns, it is possible to employ multimodal
features that combine text and visual features to detect potentially harmful scripts and
Sensors 2024, 24, 4152 3 of 31
risky behaviors such as memory or resource usage. This unique multimodal technique
accurately identifies a wide range of intrusion threats. This paper introduces an innovative
approach for detecting IoT-based NIDS involving analyzing and characterizing network-
based multimodal features. A crawler is developed to gather network flows from PCAPs
for in-depth study. Subsequently, word2vec is used to extract trained vocabulary features.
The network-based byte stream is then converted into an image. We develop a powerful
NIDS by integrating text-based and visual features. Our findings show that these two
features enhance one another, resulting in a higher intrusion detection rate. The proposed
methodology establishes a strong foundation by incorporating anomaly detection and
signature-based approaches. Although the current focus is on signature-based detection
approaches, including word2vec transfer learning to extract features from harmful scripts,
the system can be readily adapted to detect anomalies. This adaptability provides the
ability to detect network anomalies. Anomaly detection algorithms can be developed using
dynamic feature extraction techniques and enhanced streaming data processing capabilities.
This development can enable the IDS to detect abnormal network activity in real-time,
thereby improving signature-based detection. Consequently, it can offer a comprehensive
defense against emergent threats. The main contributions of this paper are as follows:
• We proposed custom datasets crawled from PCAPs in the CIC-IoT 2022 and 2023
datasets. Numerous attacks are assembled, including camera-based flood, DDoS,
RTSP brute force, etc. Furthermore, PCAPs are mined for various camera-based
attacks to analyze abnormal visual surveillance behavior.
• The optimization approach, implemented within the Spark framework, efficiently
extracts insights from huge datasets. Comparative analysis of different optimization
strategies is used to determine how big data behaves and to improve the workings
of the NIDS. The word2vec transfer learning approach extracts trained features from
dangerous scripts while minimizing data transmission overhead. This strategy uses
semantic anchors to focus on specific network attacks involving malicious scripts.
• We design a method that converts the network bytes to an image to analyze the visual
features. A malware-to-image conversion algorithm is developed that can transform
the byte stream into a grayscale image. The texture features are then extracted from
visuals using an attention-based Residual Network (ResNet)-trained model. The text
and texture features are then combined for effective IoT-based NIDS.
• A game theory-based method is designed to validate the performance of the proposed
method, supporting the use of the Nash equilibrium and mathematical formulations
to develop a reliable and trustworthy IoT-based IDS system.
The remainder of the paper is organized as follows: Section 2 describes the relevant
and recent literature; Section 3 describes the proposed approach; Section 4 provides an
extensive experimental analysis; Section 5 presents the IDS game theory-based study; and
Section 6 concludes the work.
2. Related Work
Numerous research studies have been proposed on IoT security using IDS customized
to the IoT ecosystem’s characteristics. Stephen et al. [13] designed an approach renowned
for its mobility, composite structure, and centralized design. This method is intended to
detect Hello Flood and Sybil attacks on IoT networks that use the Low Power and Lossy
Networks (RPL) protocol for routing. Their method uses detection measures, including
the number of packets received and delivered, to verify the intrusion ratio (IR) via the
IDS agent. Raza et al. [14] designed the SVELTE method, a real-time IDS for the Internet
of Things. It includes a 6LoWPAN Mapper, an intrusion detection module, and a small
firewall. The method examines traced data to detect unauthorized network access and has
been particularly tested for its capacity to identify faked or changed information, sinkholes,
and selective forwarding threats. Shreenivas et al. [15] made SVELTE better by adding an
intrusion detection module with an Expected Transmission (ETX) measure to find bad activ-
ities happening on the network. They also suggested employing a geographical indicator
Sensors 2024, 24, 4152 4 of 31
to identify hostile nodes engaged in ETX-based network attacks. Combining the EXT and
rank-based methods revealed a rise in the overall True Positive (TP) rate. Pongle et al. [16]
proposed an IDS with centralized and distributed architectures, which they constructed
and tested using simulated scenarios and networks. This technique is specifically designed
to detect routing attacks such as wormholes. Jun et al. [17] presented an event-processing-
based IDS for IoT. For attack detection, this specification-based system employs complicated
event-processing algorithms. It collects IoT data, extracts events, and matches them to rule
pattern repositories to detect security instances. While it is more efficient than standard IDS,
it does place a large demand on the CPU. Summerville et al. [18] designed an IDS for IoT
by combining a deep packet inspection approach with a bit-pattern strategy. With feature
selection using overlapping n-grams, network payloads are considered as bit-patterns.
The bit pattern and n-grams are said to be matched when corresponding bits line up in
every location. The false positive rate of the system decreased substantially when it was
evaluated via the execution of four assaults [19]. Ioulianou et al. [20] proposed a hybrid
lightweight signature-based IDS to counter two DoS attacks: Hello flooding and version
number alteration. Despite the excellent findings, it is vital to highlight that their system
was tested in a simulated environment using Cooja.
Recent research has shown increasing interest in using big data gathered from network
traffic to find different patterns and classify harmful activity. The proliferation of big
data has facilitated the advancement of more precise IDS by implementing ML and DL
methodologies [21,22]. Gupta et al. [23] used Spark streaming to perform real-time network
analysis. Their research revealed several streaming analytics issues, notably spotting port
scans, tracking mirrored breaches, and analyzing application efficiency. Using programmed
switches such as OpenFlow switches allowed for the collection of 95% of query-related
activity. Mustapha et al. [24] used Apache Spark to compare the efficacy of four popu-
lar classification methods for detecting intrusions in network traffic: SVM, NB, DT, and
RF. The UNSW-NB15 dataset was used to assess detection accuracy and forecast time.
Samed et al. [5] developed an IDS using a hybrid deep learning (HDL) network with CNN
and LSTM models. To reduce the influence of data imbalance on system performance, they
used the Synthetic Minority Oversampling Technique (SMOTE) and Tomek-Links (STL)
sampling methods. The PySpark application within Apache Spark showed a remarkable
classification accuracy of 99.17%. A statistical IDS was put forth by Kamran et al. [25]
to detect fraudulent information attacks and additional forms of intrusion. They inves-
tigated the consequences of numerous parameters using simulation and analysis under
various traffic situations, presenting their modeling findings graphically and quantitatively.
A cooperative information exchange application layer (IDS) was used to compare and
evaluate the methodology. The CIDDS-001, CICIDS17, and CICDDoS19 datasets were
used by Pontes et al. [26] to design an energy-based method for finding intrusion attacks.
Traffic flows were classified as benign or detrimental according to their energy values. To
accomplish this, the authors utilized real-world data and executed Mixed Integer Linear
Programming (MILP). Their method reduced the Total Cost of Ownership (TCO) by 35%
compared to server count and 43% compared to turnaround time.
Business applications frequently produce large amounts of log data to track operational
conditions and network utilization metrics. An IDS generally necessitates the examination
of PCAP files, which contain unstructured network data. The data may contain vital
characteristics of specific attacks, including names, references, and host details. Semantic
preprocessing of network data removes noise and extracts relevant information, potentially
relieving network-based IDS [27]. Seyyar et al. [28] suggested a web IDS solution that
uses network traffic to categorize URLs as benign or malicious. The technique uses the
BERT transformer model and CNN to categorize distinct types of attacks during the URL
analysis phase. Li et al. [29] used the word2vec model and weighted mapping of HTTP
data to design a low-dimensional feature vector for semantic extraction. The detection
accuracy, true positive rate, and false positive rate were high on datasets such as HTTP
CSIC 2010, UNSW-NB15, and Malicious URLs. Min et al. [30] advocated a statistical
Sensors 2024, 24, 4152 5 of 31
and payload-based IDS strategy using word semantics and text-CNN to extract relevant
attributes from payloads.
The work described in this paper overcomes several deficiencies in existing intrusion
detection systems. Such difficulties include scalability issues with huge datasets, which
create processing delays and limited feature representation based only on text. Other
limitations include insufficient threat detection due to a lack of connectivity between text
and image data, difficulty identifying complex assaults in diverse IoT environments, and
reliance on manual rules, limiting adaptability. The suggested solution addresses these
issues by combining text and picture data. In addition, transfer learning improves accuracy
and adaptability in IoT security and optimizes algorithms for greater efficiency. This
integrated strategy improves the system’s ability to respond quickly to emerging threats
while capturing a wider range of attack patterns, ultimately increasing overall efficacy.
3. Proposed Method
Figure 2 depicts the overall process of extracting multimodal features and detecting
attacks. We use multimodal features combining text and visual elements to detect malicious
attacks. Textual features are collected from network traffic using Spark-based optimization
algorithms and a transfer learning approach. The network bytes are transformed into
images to extract textural information. An algorithm is created to handle this conversion,
and texture features are extracted by tuning the attention-based ResNet. The trained text
and texture features are integrated and used as multimodal features to classify different
attacks. These novel multimodal features correctly identify a wide range of intrusion threats.
Figure 2. Proposed framework for enhancing IoT security using multimodal data representation and
big data.
transfer learning. Nevertheless, this type of data is perturbed and might not possess
significant interpretations that could aid in detecting diverse attacks. As a result, all chaotic
flow events are eliminated from this data. We accomplished this by developing a semantic
crawler that processes all flow activities and transforms them into substantial trends. The
subsequent steps provide additional illustrations of the crawling procedure:
• To avoid redundancy, eliminate repetitive features in sequence within input sets.
• The dataset excludes brief sequences that lack sufficient data to determine relevant
network behavior.
• Sequence length homogeneity is critical for effective IDS, as varying lengths might
confuse neural network models. To achieve balance, this technique employs a prede-
termined sequence length, designated as L. Patterns longer than L retain their initial
L elements, while shorter patterns are equalized using zero-padding. The complete
architecture is shown in Figure 2.
8 End for
9 Finish
Figure 3. A chunk of grayscale images (128 × 128) extracted from network traffic.
In neural networks, attention frequently refers to strategies that allow the model to
focus on specific portions of the input sequence or image while making predictions. A
Residual Network (ResNet) is a deep neural network that uses residual blocks or skip
connections to address the vanishing gradient issue [31]. Figure 4 shows how we con-
structed the attention-based ResNet to extract image texture information. This network’s
attention-aware characteristics are produced via several attention modules. As the depth
of the layers grows, these attention-aware characteristics evolve gradually throughout
modules. The extracted images are used as input to the ResNet blocks. The details are
as follows:
(a) Hierarchical network composition: Several attention modules are assembled in a
layered framework to develop residual attention networks. This stacking structure
is the basic implementation of the mixed attention mechanism, which allows for
integrating multiple types of attention into distinct modules.
(b) Attention Residual Learning: A significant performance drop would be the outcome
of directly stacking the attention modules; to address this, we present a method for
learning attention residuals that can enhance the performance of a residual attention
network with several layers.
Adding self-attention to the ResNet increases its ability to store long-term dependen-
cies, assess global context, and focus on key features. This combination improves data
transmission, enables feature learning at several sizes, and works across multiple tasks,
making it an effective solution for various applications.
Sensors 2024, 24, 4152 8 of 31
frequent access and computation. It is critical to note that storing data in memory can
result in insufficient memory or out-of-memory (OOM) issues when memory space is
limited or the dataset is very large. In such instances, using disk caching or studying
other optimization methods to improve data processing throughput becomes crucial.
(c) Serialization: Spark utilizes the Kryo and Java serialization mechanisms. Despite
the widespread use of Java serialization, the default application of this strategy is
inefficient, leading to large amounts of serialized data, expensive disk storage, and
network transmission expenses. In contrast, Kryo serialization provides an efficient
binary alternative. In addition to reducing the quantity of data, it speeds up the
serialization and deserialization procedures and lowers disk storage and network
transmission costs. As a result, Kryo serialization is suggested in scenarios requiring
high performance.
(d) Data storage: Spark supports various formats, including CSV, JSON, XML, PARQUET,
ORC, AVRO, and others. Choosing parquet files with fast compression can improve
Spark jobs, resulting in better performance. Parquet files internal to Spark contain
metadata, including schemas, data types, and additional pertinent information. As a
result, implementing parquet files can significantly accelerate processing and improve
the overall efficacy of information management.
(e) API selection: The three types of Spark APIs are DataFrame, DataSet, and RDD. RDD
is used in reduced-level processes, which have a restricted number of optimization
strategies. DataFrame is generally the best option, as it employs the catalyst optimizer,
which produces a query plan that enhances performance. DataFrame also has minimal
trash collection overhead. DataSets offers high-type safety during serialization by
integrating the encoder and using Tungsten as a binary serializer. The amount of
data and the techniques employed influence how rapidly RDD, DataFrame, and
DataSet in Spark process information. DataFrame and DataSet frequently beat RDD
by employing Spark SQL’s optimizer and code generator. In addition to binary
serialization and deserialization, they provide columnar storage. Furthermore, both
DataFrame and DataSet enable rigorous typing, which allows for the examination and
avoidance of type problems during compilation. DataSet requires more code and type
declarations than DataFrame but has faster processing as it employs the Tungsten
engine, a high-performance in-memory management and serialization engine. As a
result, the right API must be chosen based on specific parameters, such as data types
and processing needs.
The complexity of high-dimensional network traffic data makes it difficult to use big
data for intelligent IDS effectively. Selecting useful characteristics while lowering dimen-
sionality remains a challenge. Spark’s DataFrame, memory caching, Kryo serialization,
PARQUET file storage, and adjustable partition numbers based on node count dramatically
improve processing performance and data efficiency, especially with huge datasets. Several
important features of the IDS make proactive threat identification and response possible
in real-time. Utilizing optimization strategies within the Spark framework is one of these
methods for efficiently processing streaming data. Furthermore, the word2vec transfer
learning approach enables real-time adaptive feature extraction from harmful scripts. The
system can rapidly identify potential risks by analyzing irregular visual surveillance behav-
ior by scanning PCAPs for camera-based assaults. The IDS is ensured to be adaptable and
resilient in dynamic environments by utilizing a formal validation method based on game
theory. This method provides an exhaustive framework for assessing the effectiveness of
proactive strategies.
Sensors 2024, 24, 4152 11 of 31
Scalability is critical to our system’s response to the increasing hazard of flood at-
tacks on IoT applications. By integrating multimodal big data representation with transfer
learning, our system can dynamically adapt to various attacks. Spark-based optimization
ensures that huge datasets are processed smoothly. We use transfer learning to extract
semantically rich features, improving adaptation to evolving threats. Furthermore, we em-
ploy innovative approaches such as transforming network data into images and applying
ResNet to extract texture properties. This combination makes it possible to classify attacks
accurately. To demonstrate scalability, we extensively tested several typical IoT datasets,
such as the CIC-IoT 2022 and 2023 datasets. In addition, by including validation based
on game theory, we ensure that our system is resilient and scalable, making it capable of
efficiently handling future issues.
4. Experimental Results
4.1. Datasets
We used three standard datasets to investigate the efficiency of the proposed method,
namely, CIC-IoT 2022 [35], CIC-IoT 2023 [36], and Edge-IIoTset [37]. Canadian Institute
for Cybersecurity datasets were used in these experiments. IoT devices were set up and
used to examine network traffic behavior. The CIC-IoT 2022 dataset is generated us-
ing Wireshark and dumpcap in six distinct types of experiments. Dumpcap is used for
semi-automated tests, whereas Wireshark is utilized for hands-on testing. The tests are
divided into six categories: power, idle states, interactions, scenarios, active states, and
attacks. For this investigation, we collected network flows associated with eleven differ-
ent types of flood attacks, including Amcrest, Arlo Basestation Camera, ArloQ Camera,
Borun Camera, DLink Camera, HeimVision Camera, Home Eye Camera, Luohe Camera,
Nest Camera, Netatmo Camera, and SimCam, during device activation and interaction.
The CIC-IoT 2023 dataset operates in real-time and serves as a benchmark and dataset
specifically selected for evaluating extensive attacks in the IoT domain. Its main goal is
to introduce a unique and comprehensive dataset on IoT threats for use in developing
security analytics applications in real IoT operations. To this end, it implements 33 at-
tacks on 105 IoT devices. The attacks are classified into seven types: DDoS, DoS, Recon,
Web-based, Brute Force, Spoofing, and Mirai. We organized ten unique DDoS attacks
to conduct the experiment: SYN_Flood, TCP_Flood, SynonymousIP_Flood, UDP_Flood,
ICMP_Flood, PSHACK_Flood, RSTFINFlood, HTTP_Flood, ACK_Fragmentation, and
ICMP_Fragmentation. Finally, the Edge-IIoTset dataset contains data from more than
ten IoT devices, including inexpensive digital temperature and humidity sensors. It in-
cludes 14 DoS/DDoS, information gathering, man-in-the-middle, injection, and malware
attacks on IoT and IIoT communication protocols. The dataset comprises alarms, system
resources, logs, and network traffic, with 61 significantly related characteristics out of 1176
total features.
FP
Recall = (7)
(FP + TN)
TP
Precision = (8)
(TP + FP)
Sensors 2024, 24, 4152 13 of 31
(2 * TP)
F1-score = (9)
(2TP + FP + FN)
(TP + TN)
Accuracy = (10)
(TP + TN + FP + FN)
Figure 6. Accuracy and loss dynamic curves on the CIC-IoT 2022 dataset.
Sensors 2024, 24, 4152 15 of 31
Figure 7. Accuracy and loss dynamic curves on the CIC-IoT 2023 dataset.
Sensors 2024, 24, 4152 16 of 31
Table 4 displays the performance metrics obtained with the CNN-LSTM model on
the CIC-IoT 2023 dataset. A total of ten DDoS attacks were examined. SYN_Flood has a
precision of 75%, a recall of 97%, and an f1-score of 85%. Similarly, TCP_Flood has 100%
performance across all three performance metrics. SynonymousIP_Flood has the lowest
precision (96%), recall (69%), and f1-score (80%). Overall, the performance metrics range
from 69% to 100%. Table 5 shows the performance metrics obtained by the CNN-RNN
model on the CIC-IoT 2023 dataset. SynonymousIP_Flood has the worst performance, with
a precision of 96%, recall of 64%, and f1-score of 77%. SYN_Flood comes second with a
precision of 73%, recall of 98%, and f1-score of 83%. Overall, the CNN-RNN model delivers
performance metrics ranging from 64% to 100% for all forms of DDoS attacks. Table 6
compares the performance of the CNN-LSTM, CNN-RNN, and CNN-GRU models for
average values of the performance indicators. This study was based on both two datasets.
The two datasets demonstrate that the CNN-LSTM model delivers the best performance
for detecting various intrusions. For instance, on the CIC-IoT 2022 dataset, the CNN-
LSTM model has an average precision of 98.1%, recall of 98.4%, f1-score of 97.9%, and
classification accuracy of 98.2%. The CNN-LSTM model achieves a 96.4% classification
accuracy, 97% precision, 96.1% recall, and 96.1% f1-score on the CIC-IoT 2023 dataset. These
results demonstrate that the CNN-LSTM model has the highest performance, whereas
the CNN-GRU model has the lowest and the performance of the CNN-RNN model lies
in the middle. Table 6 shows the classification performance of the CNN-LSTM model on
the Edge-IIoT dataset, which includes 14 different IoT attacks. XSS, Ransomware, and OS
Fingerprinting have the highest classification, while MITM has the lowest. Considering
the variety of attacks in this dataset, the overall classification performance is impressive at
96.1% for 14 classes. Table 7 shows the performance comparison for the three datasets.
The confusion matrices for the three deep learning models using both datasets are
shown in Figure 8. These matrices examine the correct and incorrect classifications at
the class level. This evaluation can comprehensively evaluate how well our suggested
approach worked for the specified class. The diagonal values represent the correct classifi-
cation, while the off-diagonal values represent the incorrect classification. The CIC-IoT 2022
dataset contains eleven types of attacks, whereas the CIC-IoT 2023 dataset contains ten.
The CNN-LSTM model performs well on the CIC-IoT 2022 dataset; most classes have 100%
classification, such as Amcrest, ArloQ Camera, DLink Camera, etc. However, a few classes
exhibit slight misclassifications; for example, the Arlo Basestation Camera has 1% incorrect
classification, while the Borun Camera has 4%. Amcrest has the lowest performance when
using the CNN-RNN model, with 73% correctly classified and 27% incorrectly classified.
Using the CNN-GRU model, the lowest performance for the HeimVision Camera is 70%
correct classification and 30% error. The CNN-LSTM model delivers the best classification
Sensors 2024, 24, 4152 18 of 31
results for all sorts of attacks on the CIC-IoT 2022 dataset. However, CNN-LSTM delivers
the lowest classification on the CIC-IoT 2023 dataset for SynonymousIP_Flood, with 69%
correct and 31% incorrect, while the majority have 100% or more than 97% correct classifi-
cation. The CNN-RNN model’s classification is the next best, while the CNN-GRU model’s
classification is the worst.
Figure 8. Confusion matrices for the CIC-IoT 2022 and CIC-IoT 2023 datasets.
contributes to a better understanding of how attackers interact with and respond to secu-
rity mechanisms [39,40]. This methodology helps to identify optimal processes, analyze
equilibrium conditions, and allocate resources more efficiently. Analyzing dynamic and
sequential games allows defenders to anticipate attackers’ plans while designing adaptable
and cost-effective responses. Incorporating game theory into NIDS increases risk detection,
resulting in more innovative and reliable security solutions. Furthermore, researchers have
used the theoretical framework of the Nash equilibrium [41] to identify optimal techniques
for distributing resources in intrusion detection [42,43]. In addition, evolutionary game
theory has been used to analyze the concurrent development of methods employed by
attackers and defenders in IoT systems [44,45]. These significant findings emphasize the
significance of game theory as a vital tool in advancing IoT-based IDS. Such studies provide
new perspectives on security analysis, strategy optimization, and decision-making in a
constantly changing field of cyber threats [46,47].
Figure 9. A representation of the cyclic effect of players, strategies, actions, and payoffs on the Nash
equilibrium.
Sensors 2024, 24, 4152 21 of 31
The components of the proposed game-theoretic IDS for IoT security model can be
defined as follows:
1. In the context of IoT security, the set of players, denoted as P = {P1 , P2 , · · · , Pm }
(m ≥ 2) in this game-theoretic model, consists of defenders and attackers. In IoT
systems, defenders aim to prevent unauthorized access and protect system stability.
At the same time, attackers, driven by objectives such as data pilferage or service
disruption, persistently threaten these networks.
2. The concept of ‘activity’ is essential in the game-theoretical analysis of IDS, as it
encompasses the strategic decisions made by players, whether attackers or defenders,
to optimize their profits. The game’s rules lie in the players’ ability to decide between
interacting or dodging binary options. Attackers need to decide whether or not to
initiate attacks; at the same time, defenders must choose whether to execute their
strategy or adopt a passive approach. This strategic duality defines the game’s
interactive essence. Within a formal setting, we shall label P to indicate a player with a
collection of non-trivial feasible options. The set A is defined as A = {A1 , A2 , . . . , Am },
where every component Pn represents a distinct action that the player can execute.
The intrinsic depth of a set A contributes to the game’s strategic complexity by offering
player P a multitude of possibilities to achieve the most advantageous outcome.
3. In an academic context, a collection of strategies in game theory is represented as
S = {S1 , S2 , . . . , Sm }. These methods include a sequence of planned actions formu-
lated by subjects while considering previous outcomes and their rewards. The efficacy
of the attacker’s strategy relies on evading detection while transmitting malicious data
packets. On the other hand, the defender’s strategy focuses on identifying unusual
changes in entropy and efficiently coordinating defensive actions while minimizing
energy consumption.
4. The payoffs are decided according to each participant’s chosen strategies. The col-
lection of payoffs consisting of m payoffs for matching strategies can be denoted as
Y = {Y1 , Y2 , . . . , Ym }. The payoffs are determined by considering the advantages of a
successful defense or attack, the cost associated with deploying defensive methods,
and the costs borne by attackers. The provided evaluation approach clearly outlines
the effectiveness of each defensive strategy against each attack strategy, indicating
how well each defense can neutralize a certain sort of attack.
5. Upon analyzing the methods used by attackers and defenders and the resulting
benefits, it is crucial to understand how these benefits interact with the strategic
structure of the game. Such interactions can be defined by the concept of the Nash
Equilibrium (NE) [41]. The NE is important in game theory, as it represents a scenario
in a non-cooperative game where no player can improve their result by independently
altering their strategy given that the other players’ strategies remain unchanged. The
mathematical representation of the NE is as follows:
• Consider a game with m ≥ 2 players, denoted as P, each possessing a distinct
set of tactics, say, SP . Let sr denote the strategy selected by player Pr and let s−r
indicate the strategies chosen by the remaining players. The utility function for
player Pr is represented as UP (sr , s−r ).
A Nash Equilibrium is defined as a strategy profile s1∗ , s2∗ , . . . , s∗m in which the
•
condition stated below is satisfied for every player Pr :
Figure 10. A representation of the proposed strategies of the defender and attacker.
The representation of the strategy space for both the defender (SD ) and attacker (S A )
can be identified by the rows and columns in a 3 × 3 matrix M, as follows.
q11 q12 q13 SD1 , S A1 SD1 , S A2 SD1 , S A3
M = q21 q22 q23 = SD2 , S A1 SD2 , S A2 SD2 , S A3
q31 q32 q33 SD3 , S A1 SD3 , S A2 SD3 , S A3
In the above matrix M, the element q11 denotes the rate-based detection of a DDoS attack by
an IDS against volumetric DDoS attacks executed by the attacker. This particular situation
entails the utilization of volumetric DDoS attacks by the attacker, which is already known to
the defender. Simultaneously, the defense employs rate-based DDoS monitoring to identify
and detect attacks efficiently. The calculation of the IDS reward depends on the equilibrium
between energy use and the benefits obtained from efficient monitoring, as defined in the
following equation:
U11 ( D ) = GSD (t) − ERb (t). (12)
The attacker experiences a loss of the resources utilized due to the attack’s lack of success.
Equation (13) provides the calculation for the attacker’s payout:
Description Symbol
Energy consumed by Rate-based DDoS IDS ERb
Energy consumed by Anomaly-based IDS E Ab
Energy consumed by Heuristic Network Behavior IDS EHNB
Gain for successfully detecting GSD
Value of the assets under attack VAA
Resource consumption by attacker RCA
Gain for successfully attacking GSA
Waiting time cost for attacker WCA
Detection rate of IDS for Rate-based DDoS α
Detection rate of IDS for Anomaly-based β
Detection rate of Heuristic Network Behavior γ
False Positive rate of Defender µ
Sensors 2024, 24, 4152 23 of 31
The utility matrices (MD ) and (M A ) in our cybersecurity game model represent the
strategic interaction between the defender and attacker. Positive values GSD and GSA
represent the advantages obtained from successful defense and attack, respectively; on
the other hand, the costs ERb , E Ab , EHNB , RCA , and WCA have a negative impact on utility,
indicating a balance between possible benefits and their associated costs. To calculate the
game matrix, we can use the provided strategies and payoffs in the following manner:
• Matrix for the defender’s payoff (MD )
d11 d12 d13 GSD (t) − ERb (t) − ERb (t) − ERb (t) − VAA (t)
MD = d21 d22 d23 = GSD (t) − E Ab (t) − E Ab (t) GSD (t) − E Ab (t)
d31 d32 d33 − EHNB (t) − VAA (t) − EHNB (t) GSD (t) − EHNB (t)
We have used the scribing approach [47] to identify the NE in analyzing the proposed
game-theoretic framework. For this, it is necessary to contrast the optimal function of
matrix M A with the equivalent element in matrix MD , then identify a solution based on the
pure strategies. The attacker’s maximal profits are represented by a13 and a31 ; on the other
hand, the highest possible profits for the defense are denoted as d11 , d21 , d23 , and d33 . From
the above model, we can draw two important conclusions.
Theorem 1. There is no NE in pure strategies for the suggested game-theoretic model, which is
represented by the utility matrices MD and M A .
Proof. This conclusion can be derived from the observation that the attacker’s optimum
choices (a13 and a31 in matrix M A ) do not correspond to any of the defender’s optimal
choices (d11 , d21 , d23 , and d33 in matrix MD ). Given that an NE necessitates that each
player’s strategy be the optimum response to the other player’s approach, the absence of
concurrence between the attacker’s and defender’s optimal strategies suggests that a pure
strategy for the NE does not exist in this scenario.
Theorem 2. The attacker should consistently initiate the attack to optimize the potential gains
within the suggested game-theoretic framework.
Proof. Within the framework of the proposed model, the utility matrices M A and MD
are used to assess the attacker’s strategies by considering their prospective benefits and
expenses. The utility matrix M A reveals that actions such as a13 and a31 provide the
most benefits for the attacker, as they correspond to situations where the attacker decides
to initiate an attack. The benefits of these gains are compared with the disadvantages
of waiting or using other techniques that do not include attacking, which tend to be
comparatively less advantageous. Given that the attacker’s main goal is to optimize their
gains, it is preferable for them to use techniques that result in the maximum level of
utility. Hence, based on the configuration of the utility matrix M A and the comparative
assessments of the advantages and disadvantages linked to each strategy, it is apparent that
engaging in an attack results in the highest benefit for the attacker. Consequently, it can be
deduced that the attacker will consistently select an attack to optimize their utility.
As the detection rates of the IDS for Rate-based DDoS, Anomaly-based, and Heuristic
Network Behavior attacks are α, β, and γ, respectively, from Table 9 we have the following
payoff matrices for the defender and attacker:
h i
αGSD (t)− ERb (t)−(1−α)VAA (t) βGSD (t)− E Ab (t)−(1− β)VAA (t) − EHNB (t)−VAA (t)
M̃D = − E (t)−V (t) βG (t)− E (t)−(1− β)V (t) γG (t)− E (t)−(1−γ)V (t)
Rb AA SD Ab AA SD HNB AA
and
(1 − α) GSA (t) − RCA (t) (1 − β) GSA (t) − RCA (t) GSA (t) − RCA (t)
M̃ A = .
GSA (t) − RCA (t) (1 − β) GSA (t) − RCA (t) (1 − γ) GSA (t) − RCA (t)
To estimate the relevant payoffs for both the defender and the attacker in each of the
following scenarios (see Table 10), we use the modified 2 × 3 matrices, denoted as M̃D
and M̃ A .
Case 1: In the above context, the defender uses the rate-based DDoS and anomaly-based
methods with the corresponding probabilities η1 and (1 − η1 ). Similarly, the probabil-
ities associated with the attacker initiating a volumetric DDoS attack and an RTSP
brute-force attack are represented by the corresponding variables ϑ1 and (1 − ϑ1 ).
Considering the relevant probabilities, the following equations represent the cumula-
tive advantages for the defender (U( D )) and attacker (U( A)):
and
The following equations represent the partial derivatives of the payoffs for the defender
(U(D)) and attacker (U( A)) with respect to probabilities η1 and ϑ1 , respectively:
∂U( D )
= ϑ1 [(α − β) GSD (t) + ( E Ab (t) − ERb (t)) + ( β − α)VAA (t)]
∂η1
+(1 − ϑ1 )[( βGSD (t) − E Ab (t) + βVAA (t) − ERb (t) − VAA (t))]
and
∂U( A)
= −η1 αGSA (t) − (1 − η1 ) βGSA (t).
∂ϑ1
Sensors 2024, 24, 4152 25 of 31
Case 2: In this case, the probabilities η2 and (1 − η2 ), represent the defender’s utilization
of rate-based DDoS and heuristic network behavior techniques, respectively. Addi-
tionally, the variables ϑ2 and (1 − ϑ2 ) reflect the likelihood of the attacker attempting
an RTSP brute-force assault or a volumetric DDoS attack, respectively. The following
equations show the resulting benefits for the attacker (U( A)) and defender (U( D )):
and
∂U( D ) ∂U( A )
By solving these equations with ∂η2 = 0 and ∂ϑ2 , we obtain
Case 3: The defender uses anomaly-based and heuristic network behavior approaches,
respectively indicated by the probabilities η3 and (1 − η3 ). In addition, the variables
ϑ3 and (1 − ϑ3 ) represent the possibility of the attacker initiating a volumetric DDoS
attack or an RTSP brute-force assault, respectively. The benefits to the attacker (U( A))
and defender (U( D )) are shown in the following equations:
and
∂U( D ) ∂U( A )
Equations ∂η3 = 0 and ∂ϑ3 can be solved to obtain
Supposing that the odds for each event are identical, it can be inferred that both the
attacker and the defender will obtain the maximum advantages. Based on Cases 1–3, we
arrive at the following NE solution for the defender (SD ) and attacker (S A ):
SD = [ϑ1 , (1 − ϑ1 ), ϑ2 , (1 − ϑ2 ), ϑ3 , (1 − ϑ3 )]
−[ βGSD (t) − E Ab (t) + βVAA (t) − ERb (t) − VAA (t)]
= ,
[(α − 2β) GSD (t) + 2( E Ab (t) − ERb (t)) + (1 − 2β + α)VAA (t)]
[(α − β) GSD (t) + ( E Ab (t) − ERb (t)) + (α − β)VAA (t)]
,
[(α − 2β) GSD (t) + 2( E Ab (t) − ERb (t)) + (1 − 2β + α)VAA (t)]
−γGSD (t) + EHNB (t) − ERb (t)
,
αGSD (t) + αVAA (t) + γGSD (t)
αGSD (t) + αVAA (t) + 2γGSD (t) − EHNB (t) + ERb (t)
,
αGSD (t) + αVAA (t) + γGSD (t)
γGSD (t) − EHNB (t) − γVAA (t)
,
( β − γ) GSD (t) + ( E Ab (t) − EHNB (t)) + (γ − β)VAA (t)
βGSD (t) − γGSD (t) + E Ab (t) − EHNB (t) + γVAA (t) − βVAA (t)
(26)
( β − γ) GSD (t) + ( E Ab (t) − EHNB (t)) + (γ − β)VAA (t)
and
SA = [η1 , (1 − η1 ), η2 , (1 − η2 ), η3 , (1 − η3 )]
α − 2β 1−γ β−γ
β γ α
= , , , , , . (27)
α − β α − β α + γ α + γ β − 2γ + 1 β − 2γ + 1
Our game-theoretic approach focuses on three specific strategies for the attacker and
defender. Using this approach, we can depict a more complex model which allows us to
find the optimal strategy and Nash equilibrium (NE) within the given framework. These
methods can be employed to illustrate the complex relationships between the defender and
attacker in the proposed Network Intrusion Detection System (IDS) for Internet of Things
(IoT) security.
As IoT systems become more complicated and ubiquitous, formal validation method-
ologies are increasingly important. Formal methodologies employ mathematical models
to systematically design, construct, and validate systems to guarantee that they possess
the desired characteristics and function efficiently in various environments [48,49]. Despite
the dynamic and varied nature of the IoT environment, formal techniques can provide a
high level of assurance against safety vulnerabilities and malfunctions. Developers can
reduce the probability of significant defects in deployable systems by identifying and
fixing potential problems early in the design phase. Model checking, theorem proving,
and formal specification languages are among the methods for accomplishing this. It is
also possible to improve interactions between interconnected IoT devices by combining
game theory with formal techniques. Game theory provides a framework for modeling
strategic interactions in IoT environments with competing resources and potentially hostile
behaviors. As decisions made by one device impact the results of others, this integration
improves the potential of developing resilient and robust IoT systems by ensuring coop-
erative behavior across devices and anticipating and mitigating conflicts. Consequently,
integrating formal methodologies and game theory into the validation process can enhance
IoT systems’ robustness, security, and trustworthiness.
The computational complexity of each algorithm is shown in Table 11. The costly
terms of the proposed scheme are defined in Algorithms 1, 2, and 3.
Table 12 shows detailed comparisons with existing approaches. Verma et al. [50]
examined how machine learning and random search categorization might improve IoT
security against DoS attacks. They examined classifiers in depth to improve anomaly-
based intrusion detection systems (IDS), assessed their efficacy using key metrics and
validation procedures, and compared them on the CIDDS-001, UNSW-NB15, and NSL-
Sensors 2024, 24, 4152 27 of 31
KDD datasets. Qiu et al. [51] employed black-box access to generate a novel adversarial
assault on a DL-based NIDS in the IoT. They used model extraction with fewer data points
and a saliency map to duplicate the model to identify the most influential packet features.
Almiani et al. [52] developed a fully automated IDS to improve cloud computing security
and avoid cyberattacks. A multi-layered recurrent neural network model is used in cloud
computing environments near end users and IoT devices. Anthi et al. [53] proposed a
three-layer IDS to identify IoT network intrusions. Their system classifies and profiles IoT
device behavior detects malicious packets and identifies attack types. Granjal et al. [54]
used anomaly-based intrusion detection to detect DoS and 6LoWPAN/CoAP assaults.
According to their experiments, this method provides an effective defense against these
threats. Yang et al. [55] presented a lightweight detection technique for FL-based IoT
intrusion detection against poisoning attacks. Their scoring method filters out unusual
central server actors and their models from the global aggregate based on the size of
the dataset and the local model loss. Suge et al. [56] introduced an IDS that uses deep
learning and machine intelligence to combat security assaults in IoT networks. It compares
attack detection with LSTM and KNN algorithms regarding detection time, kappa statistics,
geometric mean, and sensitivity. The efficacy of the IDS was examined using Bot-IoT
datasets. Saeed et al. [57] presented a sophisticated security framework utilizing random
neural networks (RNNs) for intrusion detection and prevention. Using an instrumented
source code analysis technique, this method looks for out-of-bound memory accesses
during compilation. Each memory allocation generates its own set of tags and each memory
visit adds instructions to verify the tags. Our approach uses the multimodal features with
transfer learning, providing excellent classification results.
6. Conclusions
IoT applications and resources are vulnerable to flood attacks, especially DDoS attacks.
Under such attacks, the resources of the targeted device become inaccessible to authorized
users due to the enormous influx of network packets. Because of the complexity of the
system and the number of features involved, developing an intelligent IDS is extremely
difficult. Moreover, choosing adequate features that capture the distinguishing aspects of
attacks while minimizing dimensionality remains a challenge [58,59]. This paper provides
an enhanced IDS for IoT security by combining multimodal large data representation and
transfer learning. Initially, PCAPs are evaluated to extract relevant attack types and bytes.
Following this, Spark-based techniques for optimizing big data are implemented to manage
massive amounts of data. The transfer learning process is then used to generate semantically
enriched trained features. Finally, combining multimodal features, including training and
texture features, improves the classification of various cyberattacks. The CIC-IoT 2022
and CIC-IoT 2023 datasets, two well-known IoT-based datasets, were used to examine
the proposed method thoroughly. In addition, a game theory-based methodology was
utilized in the validation phase to validate the proposed method’s efficacy formally. Three
standard deep learning models, CNN-LSTM, CNN-RNN, and CNN-GRU, were designed
for the experiments. The results show that the proposed method using a CNN-LSTM model
provides the highest classification results. For instance, the proposed CNN-LSTM model
obtained 98.1% precision, 98.4% recall, 97.9% f1-score, and 98.2% accuracy on the CIC-IoT
2022 dataset. On the CIC-IoT 2023 dataset, our CNN-LSTM model obtained 97% precision,
96.1% recall, 96.1% f1-score, and 96.4% accuracy.
Nonetheless, the following limitations and challenges must be addressed before de-
ploying the proposed system in a real setting:
• Adversarial Robustness: Addressing adversarial threats is critical for IDS reliability
and efficacy in adversarial circumstances. Detecting and mitigating network traffic
data modifications can help to defend the system against sophisticated attackers.
• Privacy-preserving Techniques: An IDS with privacy-preserving methods can reduce
concerns about collecting and analyzing sensitive network traffic data. It is possible
to protect user privacy and detect threats efficiently using homomorphic encryption,
secure multiparty computation, and differential privacy.
• Cross-Domain Generalization: The system’s applicability and efficacy can be expanded
across IoT domains and surroundings. The IDS can be trained on various datasets to
adapt to deployment scenarios, including IoT applications, topologies, and protocols.
• Energy Efficiency: Improving the system’s energy efficiency is vital, especially in
IoT environments where resources are typically constrained. Energy-efficient tech-
niques for data processing, feature extraction, and model inference can minimize IDS
computing and energy requirements while maintaining performance.
• Robustness Against Evolving Threats: DDoS attacks and other cyber threats evolve as
attackers adopt new techniques and strategies. Advanced techniques can be used to
detect and address emerging threats that are not included in training datasets.
7. Future Directions
The following prospective research directions could improve the proposed system’s
performance in various ways:
• Using Explainable Artificial Intelligence (XAI) and advanced transfer learning tech-
niques with big data technology is an exciting direction. This integration uses smart
feature engineering and interpretable models. The resulting strategy promotes confi-
dence and intelligence in big data solutions by focusing on transparency, interpretabil-
ity, and usability. Future research could look into novel strategies for improving the
explainability of complex models while retaining the scalability and effectiveness of
big data analytics.
• Research strategies could be formulated for feature extraction algorithms that dy-
namically adapt to evolving network conditions and attack patterns. Reinforcement
Sensors 2024, 24, 4152 29 of 31
learning techniques that can optimize feature extraction in real time to enhance detec-
tion accuracy should be investigated.
• The integration of IDS with edge computing frameworks to enhance distributed
intrusion detection and response capabilities could allow for evaluating network-edge
lightweight feature extraction and model inference to reduce latency and bandwidth.
• Strategies can be designed for detecting and mitigating zero-day threats that exploit
previously undiscovered vulnerabilities. Analyzing behavior-based analysis and
anomaly detection techniques can help to detect novel attack behaviors without the
need to depend on previously identified signatures.
• Profiling IoT devices and analyzing their behavior can reveal security threats. Such
methods can aid in constructing detailed device profiles and detecting deviations from
usual behavior.
Author Contributions: F.U. proposed the study, simulated it, and wrote the manuscript. A.T.
contributed to formal validation and writing. S.U. wrote the algorithms. D.C. and Y.Z. reviewed and
analyzed the proposed research. All authors have read and agreed to the published version of the
manuscript.
Funding: This research received no external funding
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data that support the findings of this study are openly available in
the Canadian Institute for Cybersecurity CIC-IoT 2022 and CIC-IoT 2023 datasets at https://www.unb.
ca/cic/datasets/iotdataset-2022.htmlandhttps://www.unb.ca/cic/datasets/iotdataset-2023.html, re-
spectively. (accessed on 30 October 2023).
Conflicts of Interest: The authors declare no conflicts of interest.
References
1. Ray, P.P. A survey on Internet of Things architectures. J. King Saud Univ. Comput. Inf. Sci. 2018, 30, 291–319.
2. Minovski, D.; Åhlund, C.; Mitra, K. Modeling quality of IoT experience in autonomous vehicles. IEEE Internet Things J.
2020, 7, 3833–3849. [CrossRef]
3. Tao, F.; Cheng, J.; Qi, Q. IIHub: An industrial Internet-of-Things hub toward smart manufacturing based on cyber-physical
system. IEEE Trans. Ind. Inform. 2017, 14, 5, 2271–2280. [CrossRef]
4. Zhou, X.; Liang, W.; Li, W.; Yan, K.; Shimizu, S.; Kevin, I.; Wang, K. Hierarchical adversarial attacks against graph-neural-network-
based IoT network intrusion detection system. IEEE Internet Things J. 2021, 9, 9310–9319. [CrossRef]
5. Al, S.; Dener, M. STL-HDL: A new hybrid network intrusion detection system for imbalanced dataset on big data environment.
Comput. Secur. 2021, 110, 102435. [CrossRef]
6. Ullah, F.; Srivastava, G.; Ullah, S.; Yoshigoe, K.; Zhao, Y. NIDS-VSB: Network Intrusion Detection System for VANET using
Spark-Based Big Data Optimization and Transfer Learning. IEEE Trans. Consum. Electron. 2023, 70, 1798–1809. [CrossRef]
7. Ramkumar, M.P.; Reddy, P.V.B.; Thirukrishna, J.T.; Vidyadhari, C. Intrusion detection in big data using hybrid feature fusion and
optimization enabled deep learning based on spark architecture. Comput. Secur. 2022, 116, 102668.
8. Limkar, S.V.; Jha, R.K. A novel method for parallel indexing of real time geospatial big data generated by IoT devices. Future
Gener. Comput. Syst. 2019, 97, 433–452. [CrossRef]
9. Alsirhani, A.; Sampalli, S.; Bodorik, P. DDoS attack detection system: Utilizing classification algorithms with Apache Spark. In
Proceedings of the 9th IFIP International Conference on New Technologies, Mobility and Security (NTMS), Paris, France, 26–28
February 2018; IEEE: Piscataway, NJ, USA, 2018; pp. 1–7.
10. Moustafa, N.; Turnbull, B.; Choo, K.-K.R. An ensemble intrusion detection technique based on proposed statistical flow features
for protecting network traffic of Internet of Things. IEEE Internet Things J. 2018, 6, 4815–4830. [CrossRef]
11. Ring, M.; Wunderlich, S.; Scheuring, D.; Landes, D.; Hotho, A. A survey of network-based intrusion detection data sets. Comput.
Secur. 2019, 86, 147–167. [CrossRef]
12. Yu, L.; Dong, J.; Chen, L.; Li, M.; Xu, B.; Li, Z.; Qiao, L.; Liu, L.; Zhao, B.; Zhang, C. PBCNN: Packet bytes-based convolutional
neural network for network intrusion detection. Comput. Netw. 2021, 194, 108117. [CrossRef]
13. Stephen; R.; Arockiam, L. Intrusion detection system to detect sinkhole attack on RPL protocol in Internet of Things. Int. J. Electr.
Electron. Comput. Sci. 2017, 4, 16–20.
14. Raza, S.; Wallgren, L.; Voigt, T. SVELTE: Real-time intrusion detection in the Internet of Things. Ad Hoc Networks
2013, 11, 2661 –2674. [CrossRef]
Sensors 2024, 24, 4152 30 of 31
15. Shreenivas, D.; Raza, S.; Voigt, T. Intrusion Detection in the RPL-connected 6LoWPAN Networks. In Proceedings of the 3rd ACM
International Workshop on IoT Privacy, Trust, and Security, Abu Dhabi, United Arab Emirates, 2 April 2017; pp. 31–38.
16. Pongle, P.; Chavan, G. Real Time Intrusion and Wormhole Attack Detection in Internet of Things. Int. J. Comput. Appl.
2015, 121, 1–9. [CrossRef]
17. Jun, C.; Chi, C. Design of Complex Event-Processing IDS in Internet of Things. 2014 Sixth International Conference on Measuring
Technology and Mechatronics Automation (ICMTMA). In Proceedings of the 2014 6th International Conference on Measuring
Technology and Mechatronics Automation, Zhangjiajie, China, 10–11 January 2014; IEEE: Piscataway, NJ, USA, 2014; pp. 226–229.
18. Summerville, D.H.; Zach, K.M.; Chen, Y. Ultra-lightweight deep packet anomaly detection for Internet of Things devices. In
Proceedings of the 2015 IEEE 34th International Performance Computing and Communications Conference, Nanjing, China,
14–16 December 2015; IEEE: Piscataway, NJ, USA, 2015; pp. 1–8.
19. Santos, L.; Rabadao, C.; Goncalves, R. Intrusion detection systems in Internet of Things: A literature review. In Proceedings of the
13th Iberian Conference on Information Systems and Technologies (CISTI), Caceres, Spain, 13–16 June 2018; IEEE: Piscataway, NJ,
USA, 2018; pp. 1–7.
20. Ioulianou, P.; Vasilakis, V.; Moscholios, I.; Logothetis, M. A signature-based intrusion detection system for the Internet of Things.
In Proceedings of the Information and Communication Technology Forum, Bandung, Indonesia, 3–5 May 2018.
21. Ahmad, Z.; Khan, A.S.; Shiang, C.W.; Abdullah, J.; Ahmad, F. Network intrusion detection system: A systematic study of machine
learning and deep learning approaches. Trans. Emerg. Telecommun. Technol. 2021, 32, e4150. [CrossRef]
22. Vinayakumar, R.; Alazab, M.; Soman, K.P.; Poornachandran, P.; Al-Nemrat, A.; Venkatraman, S. Deep learning approach for
intelligent intrusion detection system. IEEE Access 2019, 7, 41525–41550. [CrossRef]
23. Gupta, A.; Birkner, R.; Canini, M.; Feamster, N.; Mac-Stoker, C.; Willinger, W. Network monitoring as a streaming analytics
problem. In Proceedings of the 15th ACM Workshop on Hot Topics in Networks, Atlanta, GA, USA, 9–10 November 2016;
pp. 106–112.
24. Belouch, M.; El Hadaj, S.; Idhammad, M. Performance evaluation of intrusion detection based on machine learning using Apache
Spark. Proc. Comput. Sci. 2018, 127, 1–6. [CrossRef]
25. Jia, Y.; Zhong, F.; Alrawais, A.; Gong, B.; Cheng, X. FlowGuard: An Intelligent Edge Defense Mechanism Against IoT DDoS
Attacks. IEEE Internet Things J. 2020, 7, 9552–9562. [CrossRef]
26. Tripathi, R.; Vignesh, S.; Tamarapalli, V.; Medhi, D. Cost Efficient Design of Fault Tolerant Geo-Distributed Data Centers. IEEE
Trans. Netw. Serv. Manag. 2017, 14, 289–301. [CrossRef]
27. Ullah, F.; Ullah, S.; Srivastava, G.; Lin, J.C.-W. IDS-INT: Intrusion detection system using transformer-based transfer learning for
imbalanced network traffic. Digit. Commun. Netw. 2024, 10, 190–204. [CrossRef]
28. Seyyar, Y.E.; Yavuz, A.G.; Ünver, H.M. Detection of web attacks using the BERT model. In Proceedings of the 30th Signal
Processing and Communications Applications Conference (SIU), Safranbolu, Turkey, 15–18 May 2022; IEEE: Piscataway, NJ, USA,
2022; pp. 1–4.
29. Li, J.; Zhang, H.; Wei, Z. The Weighted Word2vec Paragraph Vectors for Anomaly Detection Over HTTP Traffic. IEEE Access
2020, 8, 141787–141798. [CrossRef]
30. Min, E.; Long, J.; Liu, Q.; Cui, J.; Chen, W. TR-IDS: Anomaly-based intrusion detection through text-convolutional neural network
and random forest. Secur. Commun. Netw. 2018, 2018, 4943509. [CrossRef]
31. Wang, F.; Jiang, M.; Qian, C.; Yang, S.; Li, C.; Zhang, H.; Wang, X.; Tang, X. Residual Attention Network for Image Classifica-
tion. In Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA,
21–26 July 2017; pp. 3156–3164.
32. Salloum, S.; Dautov, R.; Chen, X.; Peng, P.X.; Huang, J.Z. Big data analytics on Apache Spark. Int. J. Data Sci. Anal.
2016, 1, 145–164. [CrossRef]
33. Mazumdar, S.; Seybold, D.; Kritikos, K.; Verginadis, Y. A survey on data storage and placement methodologies for cloud-big data
ecosystem. J. Big Data 2019, 6, 1–37. [CrossRef]
34. Sun, H.; Chen, M.; Weng, J.; Liu, Z.; Geng, G. Anomaly Detection for In-Vehicle Network Using CNN-LSTM With Attention
Mechanism. IEEE Trans. Veh. Technol. 2021, 70, 10880–10893. [CrossRef]
35. Dadkhah, S.; Mahdikhani, H.; Danso, P.K.; Zohourian, A.; Truong, K.A.; Ghorbani, A.A. Towards the development of a realistic
multidimensional IoT profiling dataset. In Proceedings of the 2022 19th Annual International Conference on Privacy, Security &
Trust (PST), New Brunswick, Canada, 22–24 August 2022; IEEE: Piscataway, NJ, USA, 2022; pp. 1–11.
36. Neto, E.C.P.; Dadkhah, S.; Ferreira, R.; Zohourian, A.; Lu, R.; Ghorbani, A.A. CICIoT2023: A real-time dataset and benchmark for
large-scale attacks in IoT environment. Sensors 2023, 23, 5941. [CrossRef]
37. Ferrag, M.A.; Friha, O.; Hamouda, D.; Maglaras, L.; Janicke, H. Edge-IIoTset: A new comprehensive realistic cyber security
dataset of IoT and IIoT applications for centralized and federated learning. IEEE Access 2022, 10, 40281–40306. [CrossRef]
38. Tavafoghi, H.; Ouyang, Y.; Teneketzis, D.; Wellman, M.P. Game theoretic approaches to cyber security: Challenges, results, and
open problems. In Adversarial and Uncertain Reasoning for Adaptive Cyber Defense: Control-and Game-Theoretic Approaches to Cyber
Security; Springer: Berlin/Heidelberg, Germany, 2019; pp. 29–53.
39. Agah, A.; Das, S.; Basu, K. A non-cooperative game approach for intrusion detection in sensor networks. In Proceedings of the
IEEE 60th Vehicular Technology Conference, 2004. VTC2004-Fall. 2004, Los Angeles, CA, USA, 26–29 September 2004; IEEE:
Piscataway, NJ, USA, 2004; Volume 4, pp. 2902–2906.
Sensors 2024, 24, 4152 31 of 31
40. Alpcan, T.; Basar, T. A game theoretic approach to decision and analysis in network intrusion detection. In Proceedings of the
42nd IEEE International Conference on Decision and Control (IEEE Cat. No. 03CH37475), Maui, HI, USA, 9–12 December 2003;
IEEE: Piscataway, NJ, USA, 2003; Volume 3, pp. 2595–2600.
41. Nash, J.F., Jr. Equilibrium points in n-person games. In Proceedings of the National Academy of Sciences, Washington, DC, USA,
24–26 April 1950; National Acad Sciences: Washington, DC, USA, 1950; Volume 36, pp. 48–49.
42. Liu, B.; Xu, H.; Zhou, X. Stackelberg Dynamic Game-Based Resource Allocation in Threat Defense for Internet of Things. Sensors
2018, 18, 4074. [CrossRef] [PubMed]
43. Xia, Z.; Tan, J.; Gu, K.; Jia, W. Detection resource allocation scheme for two-layer cooperative IDSs in smart grids. J. Parallel Distrib.
Comput. 2021, 147, 236–247. [CrossRef]
44. Boudko, S.; Aursand, P.; Abie, H. Evolutionary Game for Confidentiality in IoT-Enabled Smart Grids. Information 2020, 11, 582.
[CrossRef]
45. Arisdakessian, S.; Wahab, O.A.; Mourad, A.; Otrok, H.; Guizani, M. A survey on IoT intrusion detection: Federated learning,
game theory, social psychology, and explainable AI as future directions. IEEE Internet Things J. 2022, 10, 4059–4092. [CrossRef]
46. Estiri, M.; Khademzadeh, A. A game-theoretical model for intrusion detection in wireless sensor networks. In Proceedings of the
CCECE 2010, Calgary, AB, Canada, 2–5 May 2010; IEEE: Piscataway, NJ, USA, 2010; pp. 1–5.
47. Han, L.; Zhou, M.; Jia, W.; Dalil, Z.; Xu, X. Intrusion detection model of wireless sensor networks based on game theory and an
autoregressive model. Inf. Sci. 2019, 476, 491–504. [CrossRef]
48. Krichen, M. A Survey on Formal Verification and Validation Techniques for Internet of Things. Appl. Sci. 2023, 13, 8122. [CrossRef]
49. Hofer-Schmitz, K.; Stojanović, B. Towards formal verification of IoT protocols: A Review. Comput. Netw. 2020, 174, 107233.
[CrossRef]
50. Verma, A.; Ranga, V. Machine Learning Based Intrusion Detection Systems for IoT Applications. Wirel. Pers. Commun.
2020, 111, 2287–2310. [CrossRef]
51. Qiu, H.; Dong, T.; Zhang, T.; Lu, J.; Memmi, G.; Qiu, M. Adversarial Attacks Against Network Intrusion Detection in IoT Systems.
IEEE Internet Things J. 2020, 8, 10327–10335. [CrossRef]
52. Almiani, M.; AbuGhazleh, A.; Al-Rahayfeh, A.; Atiewi, S.; Razaque, A. Deep recurrent neural network for IoT intrusion detection
system. Simul. Model. Pract. Theory 2020, 101, 102031. [CrossRef]
53. Anthi, E.; Williams, L.; Słowińska, M.; Theodorakopoulos, G.; Burnap, P. A supervised intrusion detection system for smart home
IoT devices. IEEE Internet Things J. 2019, 6, 9042–9053. [CrossRef]
54. Granjal, J.; Silva, J.M.; Lourenço, N. Intrusion Detection and Prevention in CoAP Wireless Sensor Networks Using Anomaly
Detection. Sensors 2018, 18, 2445. [CrossRef] [PubMed]
55. Yang, R.; He, H.; Wang, Y.; Qu, Y.; Zhang, W. Dependable federated learning for IoT intrusion detection against poisoning attacks.
Comput. Secur. 2023, 132, 103381. [CrossRef]
56. Sugi, S.S.S.; Ratna, S.R. Investigation of machine learning techniques in intrusion detection system for IoT network. In Proceed-
ings of the 3rd International Conference on Intelligent Sustainable Systems (ICISS), Thoothukudi, India, 3–5 December 2020;
pp. 1164–1167.
57. Saeed, A.; Ahmadinia, A.; Javed, A.; Larijani, H. Intelligent Intrusion Detection in Low-Power IoTs. ACM Trans. Internet Technol.
2016, 16, 1–25. [CrossRef]
58. Ullah, F.; Alsirhani, A.; Alshahrani, M.M.; Alomari, A.; Naeem, H.; Shah, S.A. Explainable Malware Detection System Using
Transformers-Based Transfer Learning and Multi-Model Visual Representation. Sensors 2022, 22, 6766. [CrossRef] [PubMed]
59. Ullah, F.; Ullah, S.; Naeem, M.R.; Mostarda, L.; Rho, S.; Cheng, X. Cyber-Threat Detection System Using a Hybrid Approach of
Transfer Learning and Multi-Model Image Representation. Sensors 2022, 22, 5883. [CrossRef] [PubMed]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.