A Systematic Analysis of Enhancing Cyber Security Using Deep Learning for Cyber Physical Systems

In this current era, cyber-physical systems (CPSs) have gained concentrated consideration in various fields because of their emergent applications. Though the robust dependence on communication networks creates cyber-physical systems susceptible to deliberated cyber related attacks and detecting these cyber-attacks are the most challenging task. There is the interaction among the components of the cyber and physical worlds, so CPS security needs a distinct approach from past security concerns. Deep learning (DL) distributes better performance than machine learning (ML) due to its layered architecture and the efficient algorithm for extracting prominent information from training data. So, the deep learning models are taken into consideration quickly for detecting cyber-attacks in cyber physical systems. As numerous attack detection methods have been proposed by various authors for enforcing CPS security, this paper reviews and analyzes multiple ways of attack detection presented for CPS using deep learning. We will be putting the excellent potential for detecting cyber-attacks for CPS concerning deep learning modules. The admirable performance is attained partly as highly quality datasets are eagerly obtainable for the use of the public. Moreover, various challenges and research inclinations are also discussed in impending research.


I. INTRODUCTION
As there is a fast growth of technology in various communication networks and the field of computer science leads, cyber-physical systems (CPS) are rising widely in both areas, such as academia and industries.The cyber-physical systems are measured and supervised by computer-based algorithms, which are combined with networks and users.The cyber-physical systems comprise interacting network The associate editor coordinating the review of this manuscript and approving it for publication was Engang Tian .
units with physical and computational devices.The applications of CPS are making a disproportionate impact on businesses, such as in industrial sectors, healthcare, and manufacturing.
As soon as the Internet of Things (IoT) initiates, various devices with security susceptibilities are connected to cyberphysical systems, resulting in multiple attacks.It has been observed in past years that the incidents of CPS attacks have increased after the Stuxnet attack back in 2010 [1].If cyber-physical systems attacks are not perceived and reduced rapidly, they can cause massive consequences such as damage to equipment, financial losses, and public safety.So the security of CPS is one of the vital paradigms for this.But securing cyber-physical systems is also a challenging task due to its heterogeneity of components, complex interactions among cyber-physical systems, and the attack surface's complexities [2].It is observed that an intruder can randomly interrupt the dynamism of systems or encourage agitations to cyber-physical systems deprived of the security of various strategies of hardware or software, which leads to substantial social victims or the lives of humans [3], [4], [5], [6], [7], [8], [9].If cyberattacks are perceived and positioned quickly, the loss to overall systems will be measured within the acceptable time limit.Much of the existing literature on the detection of attacks is dependent on centralized architectures [10], [11], [12], [13].The attack detection schemes are usually categorized into knowledge-based and data-driven approaches [14].The residual generation method is one of the representative detection strategies in many knowledge-based systems [15], [16], [17].Usually, residual is intended by comparison of measurements of sensors and systematic model of the system.Afterward, it is equated with the static or time-variant threshold for determining whether it is an attack or not.In the case of data-driven methods, deep learning approach and heuristic algorithms are used for building models of cyber-physical systems [18], [19].If this does not follow these associations, then the attack is assumed.Apart from centralized systems, many kinds of distributed systems appear nowadays.The main challenge of designing a distributed attack detection method is monitoring cyber-physical systems without adequate information.Most cyber-physical systems lack various cyber security mechanisms, such as message authentication, which results in numerous challenges for detecting data injection attacks [20].The absence of worldwide encryption, mainly on systems engaging in dated technologies, makes it exciting to secure in contradiction of eavesdropping attacks.So, it is required to refer replay attacks.According to the report on the global cyber-physical system market and data bridge market research, the historical market and forecast CAGR is 7.8%.The traffic in global cyber-physical systems is expected to account for USD 12,356.23 million by 2028.This increase in traffic increases the burden on the CPS systems as the market increases.To overcome this problem, the researchers of both academia and industry explored this market, and as a result, the various privacy preservation methods are explored.

A. PROBLEM FORMULATION
Although there are various advantages of cyber-physical systems, these systems are susceptible to numerous cyber or physical security threats, attacks, and challenges.This occurs due to its non-homogeneous nature and dependency on sensitive and private data.This kind of planned or accidental acquaintance with these systems leads to terrible effects, which results in complex security measures.Though this leads to the undesirable overhead of networks.So the security measures of a cyber-physical system are required to formulate.Figure 1 represents the review methodology of this paper.It represents the searching process and reviewing results.The authors have read the various papers for collecting the noticeable information and deliberate the cyber physical systems, fault and failures, cyber security standards, and various challenges.

B. WHY DEEP LEARNING FOR CYBER PHYSICAL SYSTEMS (CPS)
Deep Learning (DL) [21], [22], [23] gives better results as compared to machine learning (ML).In case of passable data, deep learning models provide the best results.Even deep learning models are applied for solving cyber-physical system cybersecurity issues compared to other fields.It is also experiential that various deep learning models are anticipated in current publications for detecting cyber-physical systems' cyber-attacks.The main is not only the way to describe the difficulty of cyber-attack detections on cyber-physical systems; the main complexity arises when superimposing cyber security over cyber-physical systems [24].Various authors have not had a detailed discussion on applying deep learning methods for detecting cyber-attacks contrary to cyber-physical systems.The brief survey was given by authors [25] with a four-step framework that uses deep learning methods for detecting cyber-physical systems cyberattacks.The biggest concern nowadays is the security of CPS.Deep Learning approaches are precisely intended to handle large datasets compared to small datasets with numerous features.These methods can approximate any function as deep learning has a rich class of models.All these methods are appropriate in cyber-physical systems due to the following reasons: • Information gathered from CPSs is commonly high layered as information from countless physical sensors and cyber sensors.
• A steady development of information because of upgrades and openness to novel susceptibilities are there.
• The models should be continually refreshed with novel information to represent the drifting of the framework and further vector assaults.Deep learning models require large amounts of data for training, and CPS data is often limited and expensive to collect.Transfer learning techniques have been applied to leverage pre-trained models and overcome this challenge.Additionally, the security of CPS can also be enhanced using deep learning techniques, such as using autoencoders for intrusion detection, and generative adversarial networks (GANs) for generating adversarial examples to improve the robustness of CPS.Overall, deep learning has shown promising results in various CPS applications and is expected to play a significant role in advancing the state-of-the-art in CPS.

C. QUANTUM LEARNING WITH CPS
Quantum machine learning is an emerging field that combines quantum computing and machine learning techniques to solve complex problems.However, quantum computing technology is still in its early stages of development, and its practical applications in the field of cybersecurity and CPS are still largely theoretical.
One of the potential advantages of quantum machine learning for CPS security is its ability to perform complex calculations faster than classical computing, which could potentially speed up the detection and response to cyber attacks.However, the development of quantum machine VOLUME 12, 2024 6019 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.learning algorithms and their integration into CPS systems is still a topic of ongoing research.Quantum learning with CPS is a promising area of research, but its practical applications in the field of cybersecurity and CPS are still largely speculative, and much work is needed to develop and test quantum machine learning algorithms for real-world CPS systems.

D. DEEP LEARNING AND QUANTUM LEARNING WITH CPS
Deep learning and quantum learning are two areas of research that can have potential applications in Cyber-Physical Systems (CPS).
Deep learning involves training deep neural networks to perform complex tasks, such as image and speech recognition, natural language processing, and even autonomous decision-making.In CPS, deep learning can be used to analyze large volumes of data generated by sensors and devices in real-time, detect anomalies and potential threats, and make accurate and timely decisions to ensure the safety and security of the system.
Quantum learning, on the other hand, uses the principles of quantum mechanics to process and analyze data.It involves the use of quantum algorithms and quantum computers to solve problems that are computationally infeasible using classical computers.In CPS, quantum learning can be used to optimize the performance of the system, reduce energy consumption, and enhance security by developing quantum-resistant encryption algorithms.
While both deep learning and quantum learning have potential applications in CPS, they are still in the early stages of development and require further research to fully understand their capabilities and limitations in this domain.

E. CONTRIBUTIONS
In this paper, we undertake an extensive investigation into the application of deep learning for cyber-attack detection within cyber-physical systems (CPS).Our contributions encompass: • Here in this paper, authors have performed an exhaustive survey of contemporary methods and techniques for cyber-attack detection in CPS, harnessing the capabilities of deep learning.
• The authors have introduced a rigorous methodological framework that serves as the cornerstone of the research.This framework not only positions our work within the current landscape but also facilitates the systematic analysis and evaluation of recent developments in this domain.
• A comprehensive examination of reliability failures and security threats, specifically tailored to the various layers of CPS architecture.
• The authors have delved into the realm of solutions with meticulous attention to technical intricacies.The discussions provide in-depth insights into the implementation of security measures, considering factors such as encryption algorithms, anomaly detection thresholds, and real-time monitoring mechanisms.
• In an alignment with the core theme, the authors have engaged in a technical discourse surrounding challenges and future trends.This includes embracing novel paradigms in machine learning (ML) and deep learning (DL), devising techniques to safeguard trained DL models from adversarial attacks, advancing the construction of CPS cybersecurity datasets with a focus on data diversity and volume, and enhancing the technical rigor of model evaluation methodologies.

F. MOTIVATIONS
As soon as the intelligent computing systems introduce predictable intelligence towards the issues of cyber, so the researchers are more inclined to use intelligent computing for secure computations, as there are various challenges for detecting attacks also.The question is whether computation spectacle can help improve security concerns.The security of Cyber physical systems is a significant concern, and that's why it is mandated to study the safety of cyber physical systems.So an analysis of cyber security of the cyber physical system is required, and it is presented in this work.The taxonomy of paper is shown in Figure 3.

G. ORGANIZATION
The rest of the paper is categorized into various subsections: Section II describes the literature review of evolution of cyber physical systems and deep learning.Section III discusses the Mathematical Modeling Framework for Enhancing Cyber Security in Cyber-Physical Systems using Deep Learning.Section IV describes the fault, failures and attacks in cyber physical systems.Open issues for securing CPS are described in Section V.Challenges and Future Scope is described in Section VI.Finally paper is concluded in Section VII.

II. LITERATURE REVIEW: EVOLUTION OF CYBER PHYSICAL SYSTEMS AND DEEP LEARNING
Cyber-Physical System (CPS) is the coordination of computers with existing frameworks.The embedded computer screen, the actual control cycles, the feedback loops, and the physical approaches also influence calculations.Cyber-Physical System is near to convergence, with no association of the physical and the cyber world as a conceptual motivation.It consolidates designing representations and strategies from mechanical, ecological, typical, electrical, biomedical, compound, aeronautical, and modern designing with the models and techniques for software engineering.
As the expressions ''the internet'' and ''cyber-physical system'' originate via a similar root, ''computer science,'' which is authored by Norbert Wiener [5], an American statistician who enormously affected the advancement of control frameworks theory, would be more precise.Wiener spearheaded innovation for the programmed pointing and shooting of hostile airplane weapons.Albeit the components he utilized didn't include computerized PCs, the standards contained are like those pre-owned nowadays in computerbased criticism controller frameworks [26].The control rationale is a calculation, though one has done via simple circuits and mechanical portions, and consequently, computer science is the combination of actual cycles, analysis, and correspondence.The similitude is adept for control frameworks.
CPS is here and is mistaken for ''online protection,'' which concerns the secrecy, uprightness, and accessibility of information and has no characteristic association with actual cycles.The expression ''network protection'' along these lines is about the security of the internet and is subsequently, by implication, associated with computer science.CPS includes many testings security and protection concerns, yet these are in no way, shape, or form the main worries.
It is an innovation in that intelligence associates our actual world with our data world.Cyber Physical Systems is more essential and solid than these as it doesn't directly reference either execution draws near or precise applications like ''Industry'' in Industry 4.0.It centers as a substitute to the principal scholarly issue of adjoining the designing customs of the digital and an actual universe.One could discuss a CPS hypothesis like the ''direct frameworks hypothesis.''CPS has turned out to be a common factor in critical infrastructure because of its massive influence and commercial assistance [6].The growing reliance of crucial infrastructure on cyber-based skills has turned them susceptible to cyber-assaults like interference, auxiliary, and exclusion of data from the communique networks [7], [8], [9].' Therefore, the sanctuary of cyber-physical systems has become a perilous concern.A brief history of computer systems and cyber physical systems is illustrated in Figure 3.
Deep Learning (DL) has acquired huge consideration in previous years.It has worked on the state-of-art execution of numerous claims, remembering applications related to security for basic designs, like interruption identification, malware discovery, access control, and peculiarity recognition and orders [6].DL was presented in the late twentieth era, which was begun with the investigation of Artificial Neural Networks (ANNs).Deep Neural Networks (DNN) comprise a set of layers that gain proficiency with a progression of hidden portrayals progressively [27], [28].Higher-level descriptions contain enhanced parts of information tests that are helpful for segregation and stifle unessential highlights.Deep Learning models have worked on the cutting-edge execution in various assignments [10], [11].The summary of related works of various methods and applications are shown in Table 1.
Figure 4 delineates the general idea of cyber-physical systems and the IoT for cyber physical systems.It displays current cyber-physical systems, how elements could be separated from such frameworks, conceivable deep learning models, and the benefits of utilizing deep learning [29].Furthermore, the information gathered from existing digital frameworks is ordinarily high layered.Deep Learning models are explicitly intended to manage high layered information.Different attributes of CPS incorporate, proceed with the development of data, information float, and openness to new framework dangers.This way, it is crucial to assemble deep learning-based sanctuary models which are versatile and extendible with the information float, nonstop disclosure of new framework dangers and weaknesses [12].
This idea of ''Generalization'' is one significant issue for constructing security-based requests in cyber-physical systems as creating AI models for one situation is almost difficult to use, experiencing the same thing even in a similar setting.In this manner, it is a quintessence to zero in on speculation that deep learning models utilized in such applications are ordinarily high layered.Deep Learning models are explicitly intended to manage high-layered information.Different attributes of CPS incorporate, proceed with the development of data, information float, and openness to new framework dangers.This way, it is crucial to assemble deep learning-based security models which are versatile and extensible by the information float, nonstop disclosure of new framework dangers and weaknesses [12].This idea of ''Generalization'' is one significant issue for constructing security-based requests in cyber-physical systems as creating AI models for one situation is almost difficult to use, experiencing the same thing even in a similar setting.It is illustrative to zero in on speculation of deep learning models utilized in such applications [30], [31].

III. MATHEMATICAL MODELING BASED ON THE VULNERABILITY ASSESSMENT
The mathematical modeling framework for enhancing cyber security in Cyber-Physical Systems (CPS) using deep learning, based on the vulnerability assessment are stated below and explained in figure 5.
1) Problem Formulation: Minimize the objective function J ( ), representing the cost or vulnerability.2) System Representation: Define the CPS system as

IV. FAULT, FAILURES AND ATTACKS IN CYBER PHYSICAL SYSTEMS
A failure is an occurrence that arises when an organization diverges as of its planned performance.The failure establishes because of its inadvertent state.The origin of a fault might be internal or external.The internal faults occur due to their physical nature (such as brokerage of the component connector), and faults occur due to their design (software or hardware-related bugs) [55].Peripheral faults (External) initiate from the environmental cause like noise.
Faults may be categorized into permanent and temporary faults.However, a temporary fault occurs for short time span.It may create an error, and this may lead to perpetual failure.Similarly, Physical faults and inputs can be temporary or it can be permanent, whereas the design faults are constantly permanent.The faults which could not be analytically imitated are usually known as irregular faults.This kind of fault can be led to soft errors.The cyber-physical systems/Internet of Things (CPS/IoT) infrastructure is shown in the figure.Faults might arise at diverse layers of architecture, such as the physical layer or control layer, respectively [13].The physical layer is susceptible to interruption, direct interference, or demolition of physical items.The network layer can make the connection of devices.The monitors and controllers in the control layer are susceptible to environmental uncertainties and handling of extents and control signals [55], [56].The collection of information can be done by the information layer and is mainly vulnerable to issues related to secrecy and integrity.

A. DEPENDABLE FAULTS IN IoT
Internet of Things tends to communicate failures primarily due to its extent and heterogeneousness.Previously, traditional cyber-physical systems used to ignore or remove such failures by validating and verifying the design.Though IoT is involved in technology, it is growing in size with time.The subsequent faults might arise per cyber-physical systems layer: • Physical Layer: -Intrusion: Interference of a signal.
The quantity of associated devices and the radiation rises, affecting measurements of sensors, conveyed communications, or control indications.
• Network Layer: -Collision of Messages: In correspondence to intrusion, the quantity of interactive devices may activate communique failures such as crashes or overloading of the net-' work.-Violation of Protocol: The protocol violation occurs due to incorrect message content.
• Control Layer: -Deadline Miss: Delayed in the response of control signal.The Control loops has to survey the restraints related to timing of a cyber-physical system application.-Misusage: Sending erroneous inputs to a component • Information Layer: -Inaccessibility: Lost data instigated by a skill apprise.The things might be linked, detached, or updated in the Internet of Things.

B. SECURITY INTIMIDATIONS IN CYBER-PHYSICAL SYSTEMS
Security has been one of the biggest concerns in computer networks to identify susceptibilities and avoid malicious attacks on the devices.Whereas in cyber-physical systems, more and more susceptibilities arise in the physical area and the indeterminate behavior of the physical atmosphere.The categorization of attacks applied per cyber-physical systems layer is given below: • Physical Layer: -Information Leakage: Stealing perilous information from various devices such as private keys -Denial of Service: Manipulating various parameters for performing DoS attacks.
• Network Layer: -Jamming: Overloading the communication protocol by introducing false traffic.-Collision: Manipulation of timing, the power which leads to collision of data or violation of communication protocol.-Routing misdirects: Manipulating the routing mechanism leads to collision of data, flooding of data, and discriminating promoting of facts [57], [58].
• Control Layer: -Desynchronizing: Violating the timing or manipulation of clocks.This could lead to denial of service and leakage of information.
• Information Layer: -Eavesdropping: Stealing or sniffing of information.It is one of the biggest intimidations associated with confidentiality.Furthermore, data could also be deployed to accomplish various attacks.The potential intimidations and penalties could be stated in sanctuary intimidation models for cyberphysical systems.

C. LONG-TERM DEPENDENCE AND SECURITY INTIMIDATIONS
The Internet of things and cyber-physical systems will deviations over time, particularly when imperiled by a long operational period.Following features of the change might cause faults such as changes in the environment, functional changes, and changes in technology.The catagories of attacks implied on different CPS systems are mentioned below: • Physical Layer:At this layer there is material decay and environmental effects issues at this layer and this voilates the environment.
• Network Layer:It overloads the information by putting false traffic on the network.Due to this the communication through protocol also voilates.
• Control Layer:It disturbs the timing or manipulates the clocks this leads to the aging effects and uncertainity effects.• Information Layer:The biggest itimidation concerned with confidentiality is stealing of information.The refershing of memory, recharging is theissues relaed to this layer.

D. DEEP LEARNING IN CYBER-PHYSICAL SYSTEMS
Here we discuss how deep learning can be applied in CPS.So, the introduction of deep learning is required for how it is used in security-related applications such as CPSs.Nowadays, DL is gaining huge focus in data science to enhance performance in various applications [12].Deep Learning Algorithms contain hierarchical architectures with many layers in which higher-level features are explained in standings of lower-level features capable for the extraction of features and concepts from underlying data [14].These architectures can produce outstanding results in applications like cyber-physical systems security [12], [15].Figure 5 presents various applications of DL for CPS.The deep architectures are formed of various hidden layers [4].Deep Learning methods can represent additional abstract illustrations of information due to the multi-level architecture.

VOLUME 12, 2024
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Deep Learning models have revealed better generalization competence in many practical applications than shallow
ANNs.
There are some major fields where deep learning has been effectively applied in cyber-physical systems for security-related determinations such as detection of anomaly, detection of malware and threat hunting, susceptibility recognition, interruption detection, prevention of blackouts, assaults, and destructions in CPSs.

E. CYBER ATTACKS
In current years, there was a hike in the proportion of cyber attacks aiming cyber physical systems with distressing significances.As per recent studies [86], [97], cyber physical systems are susceptible to malicious code injection attacks [66] and code reuse attacks [76]in addition with false data injection attacks [77], zero-control data attacks [83].These kinds of attacks can lead to black out targeting cyber physical system's industrial devices and systems as shown in Table 4.

V. OPEN ISSUES FOR SECURING CPS
There are several open issues and research directions related to securing Cyber Physical Systems (CPS) using Deep Learning (DL).Some of the key areas of focus include: • Data Collection and Preparation: CPS typically generate vast amounts of data that are relevant to the security of the system [31], [55].However, collecting and preparing this data for use in DL models can be challenging, particularly when the data is highly heterogeneous and distributed across multiple sources [105].
• Model Selection and Development: There is a need to identify the most appropriate DL models for securing CPS and to develop these models so that they can be effectively applied to real-world scenarios [106].This includes choosing the right type of model, such as CNNs or RNNs, and optimizing the model's architecture and parameters to improve its performance.
• Integration with Other Security Measures: DL models need to be integrated with other sanctuary procedures to ensure that they are effective in detecting and mitigating cyber threats [107], [108].This may include integrating DL models with intrusion detection systems, firewalls, or access control systems, or incorporating additional data sources such as log data or network traffic data to improve the accuracy of the models.
• Scalability and Real-Time Processing: CPS generate huge quantities of data in real-time, which makes it challenging to use DL models to detect and answer to cyber intimidations in real-time [109], [110].There is a need for DL models that are able to scale to handle large amounts of data and that can be implemented in real-time to detect and reply to cyber intimidations in real-time.
• Explainability and Trustworthiness: One of the challenges of using DL models for security purposes is the lack of transparency and interpretability of the models.
There is a need to develop DL models that are more transparent and interpretable, so that security experts and decision-makers can understand the basis for the models' predictions and decisions [55], [56].
• Adversarial Robustness: CPS are often targeted by sophisticated cyber-attackers who use techniques such as adversarial machine learning to evade detection.There is a need for DL models that are robust to these attacks and that can continue to operate effectively even in the presence of adversarial inputs [111].
These are some of the key areas of focus for securing CPS using DL, and there is a growing body of research aimed at addressing these challenges.By developing DL models that are effective in detecting and mitigating cyber threats, and by integrating these models with other security measures, it is possible to improve the sanctuary of CPS and reduce the risk of cyber-attacks.

A. RESEARCH DIRECTIONS IN SECURING CPS USING DL
There are several open issues and research directions for securing CPS using DL techniques.Some of these include: • Development of robust and accurate DL-based intrusion detection systems for CPS: This involves the usage of DL methods such as CNNs and RNNs to detect and classify various types of cyber-attacks in CPS.
• Improving the interpretability of DL-based CPS security models: Currently, one of the main limitations of deep learning models is their absence of interpretability, making it hard to comprehend how they attain at their decisions.Research is needed to make DL models more transparent and interpretable [57].
• Anomaly detection in CPS using unsupervised DL techniques: Unsupervised DL techniques such as Autoencoders and Variational Autoencoders (VAEs) could be utilized to detect anomalies in CPS by learning the normal behavior of the system and identifying nonconformities from this normal behavior [112].
• Adversarial attacks on DL-based CPS security models: Adversarial attacks are a major concern in DL, and they pose a threat to the security of CPS systems.
Research is needed to develop defense mechanisms against these attacks and to enhance the robustness of DL-based security models [113].
• Integration of DL with other security techniques: DL-based security models can be combined with other security techniques such as firewall, detection of intrusion and anticipation systems, and encryption to create a more comprehensive and effective security system for CPS [114].
• Handling large and complex data in CPS using DL: CPS systems generate large amounts of data, and this data is often complex and unstructured.Research is needed to develop DL models that can handle this data effectively and efficiently [58], [115].

VI. CHALLENGES AND FUTURE SCOPE
The major potential fields are shown in the figure where the research areas may arise.The seven steps of the research methodology are already shown in the figure.The research literature can be improvised with our research methodology, and with the help of this, the comparative analysis can be done appropriately.The further challenges can be categorized into different directions, which are:

A. THE NEW SCENARIOS OF CYBER-PHYSICAL SYSTEM'S CYBERSECURITY
The various articles analyzed communication networks in the scenarios of cyber-physical systems [25], [103].Most of the survey papers have examined the methods of cyber-physical systems in the intelligent grids or the water treatments of plants described in the 13/27 survey papers.It is the emergent way to apply deep learning in the current industry.Deep Learning is used for detecting the faults and defects in the industrial sector of complex items [97].But these were not considered as there were no issues of cyber security covered in this.The cyber-attacks and threats usually exist in the cloud server where design models are stored.We suggest that the blockchain will be analyzed more in a broader way in collaboration with cyber-physical systems, and the variety and development of cyber-physical systems scenarios will lead to intense analysis of cyber security [116], [117], [118].
6028 VOLUME 12, 2024 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

B. IDENTIFYING CYBER ATTACKS
Most of the survey papers have analyzed the false data injection attacks.Recognizing surreptitious untruthful data injection attacks is challenging as a considerable amount of noise is being formed in the cyber-physical systems, and there is a deficiency in the mechanisms of cyber security for authenticating the devices and messages which are transmitted over the network.Some categories of false injection attacks depend on the information of invaders [119], [120].As no such advanced information is needed to initiate a denial of services attacks, individually logged packets are required for replay attacks, scanned tools for penetrating attacks, and automated tools for fuzzy attacks.However, the cyber security of cyber-physical systems is a vast area compared to cyber-attacks in contradiction to cyber-physical systems.Detection of cyber-attacks that are initiated in cyberspace and infiltrate the physical domain is a challenging task [22], [121].We assume that emergent cyber-attacks will head the defense devices, but the risk could be moderated via the data-driven approach.

C. ACCEPTING NEW PARADIGMS OF MACHINE LEARNING/DEEP LEARNING
Usually, all analyzed papers follow conventional machine learning standards, includes supervised and unsupervised learning.Around 4 papers inspected problems of regression, 3 papers are related to problems of clustering, and others are based on problems of classification.The directing usage of supervised learning reflects the value of using well-labeled data [21], [122], [123].Particularly, network packets were labeled as usual or attack traffic, and the kinds of attacks were distinguished.This dependence on labeled data is limited to the broader acceptance of machine learning or deep learning methods.We suggest that the researchers and authors use new machine learning/deep learning paradigms [124].It includes reinforcement and self-supervised learning to improve the explainability of the model.We suggest self-supervised learning flourishes in the cyber-physical system's domain as deep learning models suffer from deprived explainability.
We are expectant about predicting that the deep learning models will be further reasonable when new tools and techniques are conceived and utilized [125].

D. PROTECTING THE TRAINED DEEP LEARNING MODELS
No survey papers are measured for defending the trained deep learning models, contrary to numerous attacks.In contrast, we highlight the significance of protecting the trained deep learning models due to the computational expenditures for introducing the deep learning models [126], [127].The attackers can acquire adequate data to imitate a machine learning/deep learning model by generating many inquiries and conglomerating the outcomes.The removed data can be utilized to construct a mirroring model for the assailant to find conceivable avoidance assaults.We emphatically advocate that cyber defense be led quickly because of the ignorance of adversarial assaults in the cyber physical systems situations.

E. ADVANCING DATASETS OF CYBER PHYSICAL SYSTEMS CYBERSECURITY
Between the reviewed papers, datasets gathered in the area ruled the simulation with a proportion of 14:6.Simulated information was explored in the 2 cyber physical systems situations -shrewd matrices and vehicular organizations [128].Five papers utilizing field information picked the Smack dataset, two reports the CICIDS2017 dataset, and the other diverse datasets [129], [130], [131].
Additionally, new datasets will constantly be essential and appreciated.In a perfect world, the new datasets are publicly released field information gathered from physical testbeds.A few cyber physical system testbeds are proposed to work with recognizing cyber assaults [24].The new pattern of expanding interest in building cyber physical system testbeds may help specialists to gather superior-grade assault and defense information [132], [133].The new datasets are enormous enough to take advantage of deep learning models' power, and both new and old cyber assaults should be incorporated because cyber assaults advance rapidly.If naming information is tested, sequentially isolating the assaults from the typical traffic is a practical thought.Falsely mixing the information passages addressing assaults into a bunch of ordinary traffic records should be kept away from because the basic information increase strategy doesn't consider practicality, going after groupings, and potential connections changes.To help the headway of exploration and information, we emphatically energize more high-quality datasets increasingly to be made accessible to the local area [134], [135].

F. IMPROVISING THE EVALUATION OF MODELS
Standard execution measurements were utilized in the vast majority of the reviewed papers.Misleading up-sides were examined, precision and fault rate.This is demonstrated by authors [65] that it is fundamentally further hard to distinguish the seldom-happened assaults than the normal ones determined by the Bayesian regulations [136], [137].
Moreover, time is essential in ongoing investigations since each prepared machine learning or deep learning model's presentation will unavoidably corrupt over the long run.When the cyber develops quickly, the models prepared with old information will battle with identifying new assaults.A period rot metric was proposed in to assess a prepared model's presentation misfortune.By concentrating on the time rot, we will want to choose when the model should be retrained.We want to see future work like about cyber physical systems and cyber assaults.When top-tobottom information is created and acquired, we might hope to relieve the risk of cyber-physical systems' cyber assaults.

VII. CONCLUSION
This review gives an ongoing perspective on recognizing cyber-attacks in the cyber physical systems.In particular, an inclusive perception is obtained through analyzing the cyber-physical systems situations, recognizing cybersecurity issues, interpreting the exploration issue to the machine learning/deep learning space, developing the deep learning model, planning datasets, and lastly, assessing the model.The Cyber attacks endure as a constant and conspicuous danger to the safety and betterment of cyber-physical systems.The work shows extraordinary potential to take advantage of cyber physical system's cyber information through deep learning models as a result of their promising demonstrations.We distinguished favorable examination issues, incorporating blockchain, identifying cutting-edge, steady dangers, taking on new machine learning and deep learning standards, avoiding adversarial and attacks of model extraction, enhancing datasets, and utilizing different execution measurements.We are hopeful and sure that the examination in this field will thrive.

FIGURE 1 .
FIGURE 1. Methodology for selecting and analysing the survey.

FIGURE 2 .
FIGURE 2. Broad division of concepts discussed in the paper.

3 ) 6 ) 7 ) 8 ) 10 ) 11 ) 12 )
Threat Modeling: Define a Threat Vector T = [T 1 , T 2 , . . ., T k ] representing potential threats.4) Deep Learning Integration: Integrate deep learning models to process system information.Define the model's output as f (Input).5) Data Requirements: Specify the dataset D = {(Input 1 , Label 1 ), . . ., (Input N , Label N )} for model training.Mathematical Equations: Develop equations to quantify vulnerability levels.Vulnerability Level = g(Threat Vector, Deep Learning Output) Quantification of Vulnerabilities: Assign a vulnerability score based on the vulnerability level.Vulnerability Score = h(Vulnerability Level) Validation and Verification: Establish validation metrics to evaluate model performance.ValidationMetric = ValidationFunction (ModelOutput, GroundTruth) 9) Sensitivity Analysis: Assess model sensitivity to parameter changes.Sensitivity = ∂J ∂ Limitations and Assumptions: Clearly state any assumptions and limitations in the model.Assumption i : . . .Limitation j : . . .Comparative Analysis: Develop metrics for comparing the model against other approaches.Comparison Metric = Compare(Model, Other Models) Implications and Recommendations: Discuss the implications of the findings.Provide recommendations for practical applications.

TABLE 1 .FIGURE 5 .
FIGURE 5. Mathematical modeling framework for enhancing cyber security in cyber-physical systems using deep learning.

FIGURE 6 .
FIGURE 6. Reliability failures and sanctuary intimidations with reference to cyber physical systems layers.

FIGURE 7 .
FIGURE 7. The deep learning driven methodology for security of cyber physical systems.
Recurrent Neural Networks (RNN), and Deep Belief Networks (DBN) have been used for CPS applications.CNNs have been used for image and signal processing tasks in CPS, while RNNs have been used for time-series data analysis in CPS.DBNs have been used for fault diagnosis and anomaly detection in CPS.

TABLE 2 .
A systematic analysis of enhancing cyber security using deep learning for cyber physical systems'' vs. existing survey papers.

TABLE 3 .
Threat models for different CPS layers.

TABLE 4 .
Different CPS system with different types of anamolies.

TABLE 5 .
Real cyber physical systems attacks.