Article ID : sjrit.2024.9 | Open Access

Investigating Data Poisoning Attacks on Cyber-Physical Systems in Smart Grids (SCM) Using Edge Computing with Blockchain and DL-Driven Advanced Approach
Fateh Bahadur Kunwar | Rakesh Kumar Yadav, | Hitendra Singh
Submission Date : November 20, 2024 Publication Date : April 28, 2025
The increasing integration of smart technology in vital infrastructures, such as smart grids, raises serious concerns about their vulnerability to cyber assaults. In this study, data poisoning attacks against Cyber-Physical Systems (CPS) in smart grids are examined, with a particular emphasis on the Smart Grid Management (SCM) component. The research suggests a complex defence system that makes use of blockchain, edge computing, and an advanced deep learning (DL) strategy. Through the reduction of latency and improvement of system responsiveness, the integration of Edge Computing seeks to improve real-time processing capabilities at the network's edge. To create a distributed ledger that is safe from tampering and maintains data integrity throughout the smart grid, blockchain technology is used. Using machine learning to detect and eliminate possible risks, the DL-driven advanced strategy focuses on anomaly detection and mitigation. The study looks at possible weaknesses in SCM and evaluates how well the suggested defence method works to prevent data poisoning attacks. The research intends to highlight the significance of a comprehensive cybersecurity strategy for smart grids by offering insights into the resilience of the suggested solution in real-world settings via simulations and tests.
Currently, vast amounts of data are generated each moment as a result of the pervasive utilisation of automation, social media, and technological devices. Concurrently, there is a rise in cyberattacks, encompassing data breaches and identity theft. A variety of security measures mitigate these threats. Within the domain of the internet and digitalisation, blockchain technology represents an emerging trend that provides superior security. The current security protocols rely on centralised servers and systems. The drawbacks of the situation encompass dependence on trustworthy third parties, susceptibility to security breaches, and singular points of failure. Conversely, blockchain technology functions as a decentralised system that depends on trust among network nodes rather than on reliable external entities. The initial digital currencies that rendered blockchain technology accessible were Bitcoin, Ether, and Ripple. Blockchain technology can achieve ethics, accessibility, and privacy standards across various industrial applications, such as banking, healthcare, supply chain management, and voting, without depending on a singular authority.[1][2][3].

The fourth industrial revolution, also known as a technological revolution, has emerged from the integration of digital technology from the previous industrial revolution with the physical and biological realms.[4][5][6][7] The advanced technologies of the fourth industrial revolution, which amalgamate individual ICT with scientific methodologies, facilitate disruptive innovation. Examples of these technologies encompass genetic editing, autonomous vehicles, 3D printing, and the Internet of Things (IoT). Considering the limitations of depending solely on internal business information in a dynamic environment, integrating external knowledge may improve a company's innovation performance.[8][9]. The comprehensive strategy for managing innovation, known as "open innovation," involves systematically promoting and identifying a diverse array of internal and external sources for innovation opportunities, actively integrating this exploration with organisational resources and capabilities, and extensively utilising those opportunities through various channels.[10][11][12][13]. Artificial intelligence profoundly impacts the industry's approach to innovation. Blockchain technology could facilitate the viability and broader adoption of open innovation by enhancing intellectual property management, increasing transparency, fostering knowledge sharing, and enabling collaborative empowerment through smart contracts and open data, alongside providing novel liquidity for funding innovation, in contrast to the fragmented nature of open innovation. With appropriate data and intelligent contracts, these technological opportunities are attainable.
Artificial intelligence has significantly transformed our lifestyle. Automation is advancing at an accelerating pace across all sectors and levels[16]. Artificial Intelligence is advancing swiftly and has transformed social, economic, and political domains. Artificial intelligence has significantly transformed applications in healthcare, business, education, autonomous vehicles, tourism, social media, and agriculture. Artificial intelligence is becoming increasingly vulnerable to attacks as it is employed for more critical functions. This is due to the increasing prevalence of attacks on sectors such as healthcare, the military, and civic society.
Figure 1 illustrates the various forms of assaults on AI systems[18][19][20]. Assaults on the AI system's input, such as altering an image to modify its output, are termed input attacks. A corrupted AI system is unnecessary. There exist four categories of input attacks: physical, digital, imperceptible, and perceptible. Assaults on visible objects are termed perceivable assaults. Imperceptible assaults are characterised as attacks aimed at digital or physical targets that are not discernible to the naked eye. Digital attacks, often inconspicuous, target digital data such as documents, files, videos, and images. Physical objects are the target of physical assaults. Physical assaults are frequently direct and evident. In an AI model's deployment, the assailant's objective during a poisoning attempt is to induce intrinsic vulnerabilities that facilitate manipulation. Learning is compromised when learning algorithms in dataset poisoning identify patterns in the tainted data to construct a model. Algorithm poisoning involves the manipulation of an algorithm by identifying its vulnerabilities. Federated learning is vulnerable to such attacks since the user retains control over both the data and algorithms.[21][22][23][24][25]. A non-toxic model, trained on a thoroughly validated database, may be employed to update a hazardous model at different stages of development. Figure 2 illustrates the targets of the attacks that transpired during the AI implementation phase. Numerous critical software applications are rendered unusable in real-world scenarios due to these defects. To facilitate this, it is imperative to implement measures that reduce an organization's susceptibility to attacks and establish multiple layers of security. To avert "corrupt" outcomes, these measures would encompass identifying and rectifying any input contaminants, techniques, scenarios, or sources of data[26][27][28][29].

Figure 1 : Attacks Categories on AI
 
Figure 2 : Attack Target areas in AI implementation

Smart grids are essential for modernising the electricity infrastructure and enhancing productivity, reliability, and ecology. As electrical grids become increasingly interconnected, they become susceptible to various cyber hazards, particularly information poisoning attacks. This article examines the significance of studying data poisoning attacks on Cyber-Physical Systems (CPS) within intelligent grids and proposes a novel approach that integrates edge computing, blockchain-based technology, along with deep learning to enhance cybersecurity protocols. The essay also examines the importance of analysing data poisoning attacks on cyber-physical systems in smart grids.
An Explanation of Data Poisoning Attacks
A data poisoning attack is one in which the training data that is utilised by machine learning models is manipulated in order to reduce the effectiveness of such models. In the context of smart grids, these assaults might result in erroneous choices being made by the control systems, putting the whole power infrastructure's stability and dependability at jeopardy. It is essential to do research on the various forms of data poisoning assaults as well as their effects in order to strengthen the resilience of smart grids against increasing cyber threats.
Computing at the network's edge has a vital role to play in mitigating the risks associated with smart grids, which may be summarised as "the role of edge computing." Edge computing helps minimise latency and improves real-time processing capabilities. It does this by moving computational operations closer to the location where the data is generated. This is especially critical in identifying and mitigating data poisoning threats as quickly as possible, which will ultimately strengthen the smart grid's overall security posture.
Utilising Blockchain Technology
The blockchain technology presents a distributed and tamper-resistant ledger, which guarantees the honesty and openness of financial transactions in smart grids. The use of blockchain technology improves the data security by offering a framework that is both trustless and auditable. This not only protects against the manipulation of data but also guarantees the genuineness of the information that is sent over the smart grid infrastructure.
Approach Based on Deep Learning for More Advanced Work:
Deep learning is a subfield of artificial intelligence that has shown to be an effective method for locating unusual occurrences and intricate patterns hidden inside data. When deep learning models are included into the cybersecurity of smart grids, it becomes feasible to identify and react in real time to assaults that use data poisoning. Because these models are able to learn and adapt, they are well equipped to deal with the ever-changing and dynamic nature of cyber threats.
Synergies between Edge Computing, Blockchain, and Deep Learning:
The combination of edge computing, blockchain, and deep learning produces a powerful cybersecurity framework thanks to the synergies between these three technologies. Computing at the edge of the network makes it easier to create decentralised blockchain networks, which in turn ensures that safety precautions are spread out over the smart grid. An adaptable and self-learning defence against data poisoning assaults may be provided by deep learning models that are placed at the edge of the network. These models constantly examine the data for any abnormalities.
Security, Privacy, and Ethical Considerations:
Increasing security is of the utmost importance; nevertheless, it is of equal importance to address concerns about privacy as well as ethical issues. In order to win the confidence of the general public and prevent unexpected effects, it is essential to find the optimal compromise between stringent cybersecurity measures and individual rights to privacy.
Performance assessment and Regulatory Compliance:
Metrics for performance assessment need to be defined in order to confirm the efficacy of the suggested strategy. Regulatory compliance is also important. It is possible to run simulations and tests in order to evaluate the system's capacity to correctly identify and defend against data poisoning assaults. In addition, making sure that the planned cybersecurity measures comply with the applicable legislative frameworks guarantees that they are in accordance with the norms and rules that have already been set.
 
Experimental setup
Executing the prediction algorithm on Linux-based edge-embedded boards enabled us to replicate the edge computation environment of the smart grid. The central processing unit (CPU) of these boards typically comprises a Cortex-A7, operating at 1.2 GHz, with 256 MB of RAM and 512 MB of ROM. The data was processed using the mathematics, pandas, sklearn, and numpy libraries, and the experiments were conducted in Python. The statistical gradient descent (SGD) linear regression model, which emulates the behaviour of a dynamic model in an artificial intelligence (AI) edge environment, is the designated model for this experiment. The primary criteria for evaluation are the execution duration of the attack, loss over time (LOT), and mean squared error (MSE) loss. Utilising the compromised model to predict the test set samples facilitates the calculation of the MSE loss. Subsequently, the difference between the expected values and the actual values is calculated. The duration of an assault is defined as the interval from the initiation to the conclusion of the attack. The maximum mean squared error (MSE) divided by the total duration of an attack is one method to ascertain the level of threat (LOT). In evaluating the attack's efficiency regarding time overhead, LOT provides a more comprehensive analysis than the MSE statistic. The OptP method, a conventional offline poisoning attack strategy, has served as the foundation for numerous investigations, rendering it highly representative.
Data set
The open power dataset, which was produced from the combined cycle power plant dataset, consists of 9568 data samples acquired between 2006 and 2011. Aside from the average ambient temperature, pressure, and relative humidity, the characteristics also include the exhaust hoover and hourly temperature, with the anticipated label being the net energy production per hour. In order to simulate the appearance of real-time data streams, we followed the steps outlined in and inputted these samples in batches. The possible range of features and labels is [0, 1] because we normalised all the sample values. The features and labels maintained a constant range of values thanks to this normalisation process.
Basic parameters settings
We used 5%, 10%, 15%, and 20% poisoning rates to pollute the creek. Studies conducted in the past seldom contained poisoning rates over 20%. We utilised 0.001 as the convergence termination criterion for the algorithm. We updated the characteristic scores for the contaminated point samples align with the direction of gradient increase using a decay parameter (alpha) of 0.01.
Wang et al.[30] offer an authentication approach based on transfer learning empowered blockchains (ATLB) in order to preserve anonymity in industrial applications. By transferring locally or across regions, ATLB lowers the amount of time needed for model training and adds a transfer learning-based authentication method.
Salim et al.[31] hand over authentication (HO-Auth) strategy builds a user profile-based system for instant authorization and uses deep learning (DL) to authenticate devices. By ensuring that data from legitimate devices reaches the blockchain decentralised networks, the technique shields cloud applications against corrupt data. The suggested approach distinguishes between the malicious user and the truthful one by examining the CSI movement pattern gathered from many receivers.
Gaur et al.[32] suggest an authentication system that classifies medical data using the SVM model before storing it in a blockchain-based ledger system in the context of e-health. The authentication procedure may be automated using smart contracts, doing away with the need for a central authority.
Hammad et al.[33] combine preprocessing, feature extraction, and classification into a single unit for lower latency and cost effectiveness Using edge computing servers.
Hussain et al[34] intelligent control medical authentication system makes use of DL models for face recognition in healthcare settings, including ResNet-50, VGG-16, and the local binary patterns histogram (LBPH) method.
Hayashi and Ruggiero[35] presented an activity recognition model using the SVM algorithm is for automatic session management in smart home applications. Finally, provide a multi-phase method using deterministic trust transfer protocol (DTTP) to accomplish safe data transmission and use biometric technologies to identify intruders.
 
Scope for further research
To further examine poisoning assault and defence techniques, in further research, we want to examine increasingly complex deep learning and neural network models implemented online.
Finally, this study tackles the important problem of cyber-physical system data poisoning in smart grids, namely in the Smart Grid Management (SCM) space. A complete defence mechanism to strengthen the resilience of smart grids against growing cyber threats is offered by integrating Edge Computing, Blockchain, and an innovative strategy powered by deep learning. By integrating Edge Computing, the system's real-time data processing capability is improved, minimising latency-related threats. By using blockchain technology, data integrity and immutability are guaranteed, avoiding unwanted tampering and creating a transparent and reliable operating environment for smart grid systems. With its emphasis on anomaly detection and mitigation, the DL-driven advanced method provides an adaptable layer of defence that can change and adapt to fight new threats. The efficiency of the suggested defence mechanism is assessed by comprehensive simulations and tests, providing insightful information about its performance in many settings and practical applicability.
Findings of the study
  • The results highlight the need of a multi-layered cybersecurity approach for smart grids, taking into account the ever-changing nature of cyber threats and the significance of preventative actions to protect vital infrastructure.

The findings of this study support ongoing efforts to secure and future-proof these systems against cyber threats, ensuring the dependability and resilience of our energy infrastructure in the face of a constantly changing digital landscape. Smart grids continue to play a crucial role in modernising energy distribution systems
Figure 1 : Attacks Categories on AI
Figure 1 : Attacks Categories on AI
Figure 2 : Attack Target areas in AI implementation
Figure 2 : Attack Target areas in AI implementation
Figure 3 : MSE comparison of attacks
Figure 3 : MSE comparison of attacks
Figure 4 : LOT competition of attacks
Figure 4 : LOT competition of attacks
Figure 5 : Poisoning attacks impact on power prediction
Figure 5 : Poisoning attacks impact on power prediction
Table 1 : Online and offline poisoning attacks comparison
Table 1 : Online and offline poisoning attacks comparison
Table 2 : Three attacks Time comparison
Table 2 : Three attacks Time comparison
Table 3 : Comparative study of three attacks
Table 3 : Comparative study of three attacks
Pain Text:
Fateh Bahadur Kunwar, Rakesh Kumar Yadav, , Hitendra Singh (2025), Investigating Data Poisoning Attacks on Cyber-Physical Systems in Smart Grids (SCM) Using Edge Computing with Blockchain and DL-Driven Advanced Approach. Samvakti Journal of Research in Information Technology, 6(1) 34 - 54.