Converging AI innovation and quantum security for data-driven compliance, financial crime re-regulation

 Corresponding author: Srikumar Nayak Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution License 4.0. Converging AI innovation and quantum security for data-driven compliance, financial crime re-regulation Srikumar Nayak * Incedo Inc., Artificial Intelligence Practice, NYC, USA. World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 Publication history: Received on 26 September 2025; revised on 08 November 2025; accepted on 10 November 2025 Article DOI: https://doi.org/10.30574/wjarr.2025.28.2.3801 Abstract This study discusses the use of classical and quantum machine learning models to detect fraudulent bank transactions. Random Forest model was tested on credit card fraud detection data set and scored large percentage 99.95, AUC-ROC score/ROC is 1.0 and F1 scores are high. The most influential predictors were identified to be key features including the amount of transaction, periods between transactions, and location. In order to avoid the problem of class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was utilized, which enhanced the work of the model. Another promising study of quantum hardware scalability limits, but with multiple serious limitations, was the Quantum Support Vector Classifier (QSVC), which faces difficulty in qubit coherence and scalability challenges. These limitations did not allow the model to effectively process large data sets to better accommodate real world applications. Nevertheless, quantum models have the potential to improve the fraud detection system with developing quantum technology. This study brings out the usefulness of Random Forest in detecting fraud cases and outlines the opportunities of quantum models in the future, recommending future research, such as quantum-classical hybrid models, and the enhancement of quantum computers to meet real-time needs. Keywords: Fraud detection; Machine learning; Quantum computing; QSVC; Random Forest; SMOTE 1. Introduction Financial crime has been a threat to the stability and security of financial systems in the global arena. As there is a rapid growth in digital transactions, the financial institution has been dealing with a growing challenge of preventing fraud, money laundering, and other illegal activities. [1] Estimates indicate that global financial fraud losses have been estimated to take 33.5 billion in 2022, with a steep rise compared to 28.4 billion in 2020, and it is claimed to increase even further in the coming years. Conventional techniques in detecting financial crimes, which are largely based on rulebased systems and classical machine learning models, have not been adequate in addressing the challenges of brevity and scale of the current financial crimes. With the continuous changes in fraudulent activities, there is an increasing demand for more sophisticated technologies to deal with the problems. As part of the study, the authors are examining the potential of combining Artificial Intelligence (AI) and Quantum Computing in improving financial crime detection and compliance with regulations. AI has found extensive application in the financial sector especially in detection of fraud. The machine learning models determine a variety of trends of fraudulent activities that otherwise would be unidentified by human agents on an enormous quantity of information. The predictability and predictability of AI has been useful in detection of anomaly as well as giving predictive actions on the types of threats that may be experienced. Fraud can be predicted and otherwise accurately identified using machine learning techniques and more so, the supervised learning technique, using transaction data. But even though these systems are shown to have been effective in working efficiency, they also have World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 948 limitations in regard to scaling, flexibility, and accuracy more so when extensive inputs of complex data have to be fed into them. Quantum computing can be the means to offer the solution of these limitations, and it is transformative. Contrary to absent quantum computers, quantum computers take advantage of quantum mechanics formulation (including entanglement and superposition) to perform information processing in radically new ways. Quantum algorithms query large datasets on the same basis, in comparison to classical systems, meaning they are exponentially faster at present [2]. Such quantum algorithms can improve many financial functions, such as fraud detection, risk modeling, and portfolio optimization. Nonetheless, the continued development of quantum computing poses a major threat, especially to the security of financial systems. Eventually, quantum computers would break the encryption protocols that protect financially sensitive data; thus, there is a sense of urgency to research quantum-safe cryptography to ensure future protection. Combining AI and quantum technologies will give a great hope of making financial crimes much easier to detect without the security risks of quantum technologies. The proposed research is expected to create a hybrid framework that will incorporate the power of AI, analyzing sophisticated data, and quantum computing, which would have the power of computers and create a more efficient and secure financial crime detection system. [3] Insists on the value of postquantum cryptography, which plays a vital role in creating a quantum-safe environment, allowing for resistance to the new quantum threat to the security of financial information. This ability to introduce predictive capability of AI, as well as the power of speed of quantum computing, can offer a solution to the threats that continue to evolve on a daily basis to financial institutions. This paper also appreciates the fact that data-driven solutions are required in the detection of financial crime, as the Insights of [4] note, the ever-growing complexity of financial crimes necessitates the transition to the data-centric models capable of contemplating large volumes of transaction data in real-time. In combination with quantum computing, AI models will be able to analyze and process financial data at a new level, making the systems' tracking of fraud more accurate. The quantum machine learning (QML) algorithms integration is likely to improve fraud detection by detecting trends in large databases, which classical algorithms cannot cope with in a workable timeframe [5]. The results of the current study prove that AI search could be used in conjunction with quantum computing to improve the detection of financial crimes. Early findings show that quantum-enhanced AI models are better than conventional models in accuracy and processing speed. According to the research carried out by [6], AI models together with past transactions such as transactions have already shown great performance in tracking the fraudulent statements and the introduction of quantum computing capable of reducing the training time and enhancing the number of pattern recognition will further improve the performance of the AI models. Furthermore, the paper [7] addresses the strategy of optimizing AI model with the assistance of quantum technologies to detect new trends in fraud to discuss the perspective of the further development of fraud detection systems. There is a quantifiable answer to the problem of controlling financial crimes presented in this study. The framework prepared in this research was contrasted with the already existing financial crime detection systems and has demonstrated its ability to offer a strong, scalable, secure solution to the present-day financial institutions. The results show that quantum-enhanced AI models can be effective in reducing the number of false positives and increasing the accuracy of fraud detection and the ability of financial crime systems to adapt to emerging threats. 2. Materials and methods 2.1. Data Collection and Source This paper examines the publicly accessible data on the analysis of fraud detection in the financial market, including AI and quantum computing. The major data to be utilized in this study is the Credit Card Fraud Detection dataset on Kaggle. This dataset is one of the standard ones applied in fraud detection studies because of its practical nature and multivariate nature, comprising both discrete and numerical data. It has about 284,807 transaction records of European credit card transactions. The transaction always has a time, amount, and anonymized (pretended V1 to V28) feature, all of which offer great hints in tracing out fraudulent activity. The Credit Card Fraud Detection dataset is perfect, given that it is an imbalanced dataset in that the percentage of fraud is a small portion of the entire dataset. This class imbalance gives a learning model difficulty; however, this is modeled after fraud detection in the financial industry, where fraudulent operations are most likely few in comparison to honest transactions. The features in the dataset are 31, and all the features are anonymized to prevent sensitive data through PCA. It also consists of a binary classification target (Class = 1 in the case of fraud and Class = 0 in the case of non-fraud) that is used as the model-training and evaluation target. World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 949 Along with Kaggle data, secondary sources will be collected in the form of open-access journal articles and reports dealing with the application of AI and quantum computing as fraud detectors and prevention tools of financial crime. These sources give a theoretical background on the implementation of these technologies in the financial industry and assist in the location of the current research in the context of fraud detection techniques on the larger picture. In papers, like [8], it is shown that quantum machine learning algorithms, such as Quantum Support Vector Machines (QSVM), are used in the detection of fraudulent transactions in high-dimensional data. These works represent the prospect of quantum-enhanced algorithms to enhance the accuracy and speed of fraud detection, and this is why these findings are included in the framework of implementing the model. Moreover, the study by [9] enlightens the reader about the idea of the enhancement of AI models through the use of quantum computing in fraud detection. It has been demonstrated that the combination of quantum algorithms has an opportunity to enhance both scalability and flexibility, and this is crucial in addressing the problem of financial fraud that is becoming increasingly complex. Data Preprocessing will be a vital phase in preparing the data to be used in machine learning and the implementation of quantum algorithms. The first step is the management of missing values through the application of imputation techniques in order to have a complete dataset that can be analyzed. Characteristic preprocessing measures, like normalizing the values of features, are taken to bring the range of the numeric features to a standardized behavior to obtain the best results of the algorithms. This is necessary in situations when dealing with machine learning models such as SVM, which can be sensitive in response to input feature scale. The data is divided into test and training sets in order to assess the performance of the model. The most common split is 70-30, in which 70 percent of the data is used for training and 30 percent is used as testing, which guarantees that the model is well tested on unseen data. A novel technique could also be used to deal with the class imbalance, where examples of synthetic minority samples (like SMOTE) may be used to produce artificial examples of fraudulent transactions, so that the literature becomes more balanced. On the whole, the process of data collection is a mixture of publicly available transactional data and the information found in the peer-reviewed literature. These are sources that will not only give the raw information the algorithm developer can build, but also the theoretical basis to use the techniques of AI and quantum computers in solving financial crime detection tasks in real-life situations. 2.2. Implementation of AI and Quantum Algorithms This paper applies classical machine learning algorithms and models of quantum computing to identify unauthorized transactions in the Credit Card Fraud Detection dataset on Kaggle. The idea is to illustrate how AI and quantum imaging can be utilized in identifying monetary crime. The dataset, comprising about 284,807 transacted cases, is anonymized with features like time, amount, and PCA-transformed features (V1 to V28). The dependent variable is an indicator of a nominal (duopoly) classification of transaction: either fraudulent (1) or legitimate (0). The dataset is also especially helpful because it contains an imbalance of classes, with fraudulent transactions being a very small percentage of the whole, which reflects the resemblance to the real-life defects of fraud detection. In the case of classical machine learning, popular classes of algorithms are used, including Random Forest and Support Vector Machines (SVM), both of which are useful in tasks of classification. The ensemble learning technique known as random forest enables the creation of various decision trees throughout training and delivers the mode of the classes when used in classification issues, which aids in the detection of fraudulent transactions. SVM, however, is a supervised learning algorithm whose task is to discover the hyperplane that best separates the classes in a high-dimensional feature space. Another area of their application is where the data is not linearly separable, which is typical in fraud detection processes of SVMs. In such a situation, the linear kernel of the SVM is normally employed due to its efficiency in highdimensional spaces. The imbalance of the classes is also handled by SVMs through the elaboration of the penalty against misclassifications, which is chiefly beneficial in the case of the datasets of fraud detection, such as that utilized in the current research. These classic algorithms are coded on the scikit-learn library in Python, which is efficient in training and assessing machine learning models. To evaluate models, standard performance indicators that are accuracy, precision, recall, and F1-score are applied. Since the dataset provided has an unequal representation of classes, metrics such as the precision and the recall are of special significance because they can estimate how well the model can identify fraud without reporting too many innocent transactions as fraudulent. The F1-score, which is the combination of precision and recall, creates a balanced score that World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 950 shows the performance of the classifier. These measures are used in finding out how effective the classical AI models are in identifying fraud. Quantum Support Vectors Machines (QSVM) and Quantum Neural Networks (QNN) are being applied with the purpose to implement quantum computing in fraud detection. The models capitalize on the fact that quantum computing is capable of computing information radically differently as compared to classical models. The quantum computers are founded on the concepts of quantum mechanics such as superposition and entanglement, to assist the quantum computer in processing massive amounts of data at a much faster rate, and also in detecting complex patterns that might otherwise be hard to carve in a classical computer. The QSVM version applied in the fraud detection takes advantage of quantum as in highspace to apply a quantum algorithm to determine the best hyperplane that will be used to partition the fraudulent transactions and the legitimate transactions. It is quantum-enhanced version of the classical SVM, which can have improved and faster results because quantum computers can perform this task exponentially faster than a classical computer system. In order to do this, Qiskit library, a full open-source quantum computing infrastructure, is used to perform QSVM. The library provides the development and simulation aids of quantum circuit programming which helps to implement quantum algorithms to real-life problems, like to spot fraud. Quantum Neural Networks (QNNs) are also modeled to show that quantum computing can be used in detecting fraud. QNNs calculate data with quantum gates in multiple layers and have the benefit over less quantum-guided classical neural networks in that quantum effects like quantum entanglement are utilized. The networks can capture nonlinear relationships in the data, which is essential in complex fraud cases. The QNN can be utilized to better process, add products through quantum layers, and thus perform better ideas on task accuracy and efficiency, particularly where the data set contains complex features and is disrupted with numerous dimensions. Quantum Federated Neural Network of Financial Fraud Detection (QFNN-FFD) framework, as described in [10], is viewed as the best quantum-enhanced model. The QFNN-FFD combines quantum computing and Federated Learning (FL), a system that enables several financial institutions to train a common model without exchanging sensitive details of their transactions. The framework ensures privacy in data and enjoys better computational capability of quantum computing. The privacy of the data is also a concern that is resolved by the use of Federated Learning, as the issue of privacy is of utmost concern in the financial sector, where institutions are usually reluctant to share the details of their customers. The high percentages of precision and the strength of the QFNN-FFD model in detecting fraud are positive signs of how quantum computing can be useful in improving machine learning models to handle complex operations, including financial crimes. 2.3. Model Evaluation The quality measures of both classical and quantum models in classifying fraudulent transactions are evaluated on various standard evaluation measures: Accuracy, Precision, Recall, and AUC-ROC. These actions are important in evaluating the effectiveness of the fraud detection models especially where one of the models is denoted by an imbalanced data as noted by [11]. The accuracy is an indicator of the percentage of number of transactions that are predicted correctly. Nonetheless, Precision and Recall are also employed because of the imbalance between the classes in the detection of fraud. Precision is used to denote the ratio of fraudulent transactions that the model will detect as such to all the predicted fraudulent transactions. Higher precision means that there is a reduced number of false positives, whereas Recall is the ratio of the real fraud cases detected, and, consequently, false negatives are reduced. In order to have a balance between the two, the F1-score is employed, which is a combination of Precision and Recall. This is a crucial measure in detecting fraud because it will assist in measuring the capacity of the model used to detect fraud and incorporate both false positives and false negatives. Besides that, AUC-ROC (Area Under the Receiver Operating Characteristic Curve) is employed to determine the effectiveness of the model in distinguishing between fraud and non-fraud transactions. The larger the AUC, the better the working model, and an AUC of 1 implies perfect classification. These metrics of evaluation can be used to comprehensively evaluate the capacity of the model to deal with the problem of unbalanced data in detecting fraud. [11] Highlighted that quantum-enhanced models such as QSVM may provide a performance, especially in speed and precision, in high-dimensional data. World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 951 3. Results 3.1. Performance of Classical Model (Random Forest) The training data set used is a 10 percent sample of credit card transaction data at Kaggle, which was chosen specifically to follow a random sample of fraudulent and non-fraudulent data in Kaggle. The dataset contains qualitative and numerical variables, including the transaction sum, user actions, and purchasing places, which is why the dataset is applicable to detect fraud. Before implementing the machine learning models, a number of preprocessing was performed. In order to overcome the imbalance in the number of classes, SMOTE was used to target the number between fraudulent and non-fraudulent transactions. Besides, feature normalization was carried out to normalise the features so that all the variables are put on the same scale, and can be used as inputs to machine learning tools [12]. Figure 1 Confusion matrix for Random Forest on the 10% sample Figure 1 above indicates the Confusion Matrix of the Random Forest model, which had a perfect classification with no cases of false positive or false negative which reveals that the model had high accuracy in detecting fraudulent transactions. Random Forest, which is a popular ensemble learning algorithm, was also used to identify fraudulent transactions within the dataset. Random Forest itself works based on the notion that a number of decision trees are built, and each of hese trees is only trained using a random sample of the data and features. The average of all the trees in terms of prediction yields the ultimate prediction that offers a strong classification model that is less prone to overfitting [13]. Figure 2 ROC curve for Random Forest; AUC shown in legend World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 952 The most important KPMI in the Random Forest model are Accuracy, Precision, Recall, F1-score, and AUC-ROC. Proportion of correctly classified transactions is called accuracy. The concentration of preciseness revolves around an authentic rate of true positives (fraudulent transactions accurately recognized) among all tracked positives. Recall assesses how the model predicts all the actual fraudulent transactions and F 1 -score balances Precision and Recall into a single measure to evaluate a given model in a better way. AUC-ROC, the region under the receiver operating characteristic curve, is especially applicable in imbalanced datasets where it can be seen that it manages to distinguish between the two classes under evaluation in the model (fraudulent and non-fraudulent) [14]. Figure 2, shows the ROC curve with AUC-ROC score of 1.0, which is an indication of perfect discrimination between fraud and non-fraud transactions. The value of this ideal practice in fraud detection activities is that a high sensitivity in terms of fraud transactions is essential in fraud detection exercises. Although this outcome is the one that would be expected theoretically, one must take into account the fact that this performance could also be the result of the steps that were performed prior to it, specifically the fact that SMOTE was used to mitigate the imbalance within the classes. Such high performance is sometimes due to the overfitting especially in cases where the synthetic data is closely related to the real transactions. The Random Forest model performed very well in comparison against a simpler model, i.e. the Logistic Regression, which usually faces problems with class imbalance. Although the Logistic Regression may offer some background information to binary classification problems, it is not as efficient when identifying rare events such as fraud without thorough regression tuning or of other methods, including regularization or resampling. Figure 3 Top 15 feature importances learned by Random Forest Figure 3 shows the Top 15 feature importances that were learnt by the Random Forest model, with the transaction amount, time between transactions, location being the most important features to predict fraud. 3.2. Findings from Previous Studies Quantum Models in Fraud Detection QML has become a prospective domain of improving the current fraud detection system because quantum computing provides the power to handle more intricate datasets with greater efficiency compared to classical computing. Quantum Support Vector Classifier (QSVC): [15] Presented the use of QSVC in frauding credit card transactions. The datasets they used were fraud credit card transactions and image MNIST and Fashion-MNIST datasets. The models refined on the 4-qubit trapped-ion quantum computer, which were called QSVC, reached test poses of 70 percent on credit card dataset, 100 per cent on MNIST, and 100 per cent on Fashion-MNIST. Interestingly, the results of the QSVC models were similar to those that were derived when noiseless quantum circuit models were used, which shows the practicability of quantum models to real-life applications with noise. Quantum Kernel Methods: More recently quantum kernel methods have been combined with deep learning architecture to provide better fraud detection approaches. In may 2025, a study [16] investigated a Parliamentary talk of a hybrid system of quantum kernel methods and long short-term memory (LSTM) networks. This design is based on quantum- World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 953 enhanced feature space mappings with the use of QSVMs and the integration of the LSTM networks into the design that will capture time dependencies in transaction sequences. Results of the experiments performed with benchmark financial data showed that it was more accurate in detection, faster and more generalized than classical techniques [16]. Challenges and Scalability: The scalability of the quantum models is also a big problem even after the encouraging findings. Noise and qubit coherence times are current quantum aspects that may limit the performance of quantum models. Furthermore, there is a technical obstacle that has to be overcome to integrate quantum models with the current fraud detection frameworks, i.e., creation of effective quantum algorithms and specialized hardware is required [17]. 3.3. Classical Models in Fraud Detection Classical models of machine learning have been widely implemented and researched on fraud detection and have proven to be very robust and efficient in a wide range of real life situations. Random Forest: Random Forest algorithm is known to be one of the most efficient models that can be used to detect any fraudulent activities. In [18], a study was performed to compare several classification algorithms such as the Logistic Regression, the Random Forest and the Neural Networks to ascertain their effectiveness in detecting fraudulent activities. It was revealed that the Random Forest model is the most effective algorithm with an accuracy of 99.5% and high recall score, and it is strong in fraudulent transactions. Such a system can be put in place in live financial systems to improve on fraud prevention systems and to create safe financial transactions. Support Vector machine (SVM): SVM has been used as well in fraud detection exercises. A 2013 comparative analysis by [19] evaluated different quantum machine learning models in performing fraud in finance. The research obtained that the Quantum Support Vector Classifier model had the best performance and F1 scores of 0.98 in both the fraud and non-fraud categories. This emphasizes the possibility whereby SVM-based models be it, classical and quantum, proving successfully to classify fraudulent and legitimate transactions. Advantages and Implementations Strengths and Practical Applications: Classical models such as Random Forest and SVM are highly desirable in practical implementation as they are robust, efficient and understandable. These models have been implemented on different financial institutions, and they have been able to identify and bar fraudulent activities. They are useful in the current war against financial fraud as their capacity to process large volumes of data and adjust to changing trends in fraud suggests their utility [20]. 3.4. Evaluation of Quantum Model (QSVC) In QSVC, classical samples 𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥are embedded into quantum states ∣ 𝜙𝜙𝜙𝜙𝜙𝜙𝜙(𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥)⟩by a data-encoding circuit (feature map); classification proceeds via a kernel 𝐾𝐾𝐾𝐾𝐾𝐾𝐾𝐾(𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥, 𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥′) =∣ ⟨ 𝜙𝜙𝜙𝜙𝜙𝜙𝜙𝜙(𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥) ∣ 𝜙𝜙𝜙𝜙𝜙𝜙𝜙𝜙(𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥′) ⟩ ∣2used by a supportvector decision rule. This interpretation of the supervised quantum models as kernel methods confirms that performance and expressivity on the induced Hilbert space is controlled via the selection of embedding (feature map). As described in [21], Quantumkernel in Qiskit is a construct to create a feature map (as in, ZZFeatureMap) and execute kernel matrices using the modern primitives interface; the default implementation of this is StatevectorSampler to simulate and the QSVC to train on the resulting kernel as a classical SVC would. This [22] paper does not give empirical QSVC results because of the constraint on computation: evaluating a kernel scales quadratically with the number of samples and an execution of a circuit costs more with increasing feature-map depths, and the size of the qubit state, which makes end-to-end training on our fraud dataset prohibitively expensive with its available small datasets is hampered by the near-term availability of quantum resources and simulation expenses. A number of bottlenecks are typified in literature as apply to QSVC at scale. One, the complexity of the sample and the depth of the circuit generate interaction with the hardware noise: noise decreases the quality of the kernels, as circuits become deeper and the number of qubits increases, the margin that the SVM can achieve is lost; work around this problem tends to go to approximations and tailored tuning. Empirical works [23] on QSVC/QSVR on financial data indicate the necessity of low-rank approximations of noisy quantum kernels, and sensitive hyper-parameter optimization to get stable performance evidence that existing devices and simulators require nontrivial overheads on the quality of kernels. Third [24] practical pipelines use primitives-based simulators (e.g., State vector Sampler) or limited-qubit backends; as well as being able to prototype, such implementations ongoingly limit the size of datasets as 𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂(𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛2) circuit simulations are necessary to map 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 samples (and they also need scalable memory) [25]. Under these limitations, the QSVC is still written down and technically defined (feature maps and fidelity-based kernels) but not implemented in our results section, its analysis is still literature-based. The kernel perspective theory World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 954 encourages the notion of inclusion as a compensatory strategy and reports on the financial tasks suggest their viability on constrained regimes with approximation/mitigation, which leaves QSVC as a potential subject of future experimental study when computing and hardware will be more efficient. 3.5. General Limitations Quantum machine learning systems especially quantum kernel-based classifiers face serious practical limitations when operating in the fraud-detection context. First, such models have a computational complexity that is prohibitive. The construction and inversion (or eigen-decomposition) of an 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 Gram matrix are required by Kernel method, where n is the number of samples; worst case scenarios are 𝑂𝑂𝑂𝑂𝑂𝑂𝑂(𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛2) memory and 𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂(𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛3) computation. With quantum embedding circuits, even the overhead of estimating quantum fidelity-kernels increases exponentially with qubit count and circuit depth, making them ill posed to large data sets [26]. Second, the quantum demands acute scalability issues. Quantum kernel fidelity studies demonstrate that the value of the kernel becomes concentrated as the qubits increase in size such that the model has near identical values on all inputs preventing it from making discriminatory predictions and thus effectively avoiding its benefit. This is enhanced by largescale resampling, depth of feature-map and high-dimensionalities that are characteristic of financial fraud detection, which hinders complete deployment at scale, as indicated by [27]. Third, quantum hardware drawbacks are detrimental to practice. Current-day quantum devices are noisy intermediatescale quantum (NISQ) devices with a small number of qubits, low coherence times and highly ranging gate error rates. These limitations require overheads of error mitigation, additional decreasing effective circuit depth and data size. In the fraud detection scenario in which millions of transactions and numerous features have to be considered the hardware gap is still vast as discussed by [28]. A combination of these limitations suggests that quantum classifiers such as QSVC are currently limited in practice even though their theoretical use is promising in detecting fraud. As long as there are no quantum hardware and algorithmic scale improvements, the more realistic approach is the hybrid as well as classical-first model. 3.6. Comparison of Classical and Quantum Models. As indicated by the comparison of the classical ensemble approaches like Random Forest and quantum-kernel approaches like Quantum Support Vector Classifier (QSVC), there exist definitive trade-offs in the computing capabilities, scalability and real-world usage in a fraud detection system. Random Forest uses high numbers of parallel trained decision trees on random subsets of both data and features; calculations can be scaled basically to tree depth and forest size, and the algorithm can be well scaled to use classical hardware. It has a linear (or almost linear) scaling behavior to training data, and is used successfully with large imbalanced data including costly fraud-detection logs of transactions. By contrast, QSVC with a quantum kernel involves computing an 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 𝑥𝑥𝑥𝑥𝑥𝑥𝑥𝑥 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 Gram matrix of samples, where each kernel matrix entry involves computing (and/or simulating) a quantum circuit encoding the input and opening up the circuit to optimize costs by the square of the data size, along with circuit depth and number of qubits [29]. Scalability is another point of departure: Random Forest can provide scalability to millions of transactions and the known methods (feature selection, sampling, parallel-training) are still scalable on production systems. Quantically, QSVC is restricted by the hardware limited number of qubits available, the high number of gate errors, and the vanishing-variance limitations of fidelity kernel which diminish power of discrimination with increasing embedding dimension and data size [30]. It has also been recently demonstrated by benchmarking that fidelitybased quantum kernels can often be brought down by large datasets to classical performance equivalents, cancelling any potential quantum advantage [31]. Random Forest has demonstrated the integration of financial services in a realworld context as it is interpretable, resists unfairness, and it is efficient in its workings. In the meantime, QSVC is still mostly experimental; it is only through small groups of datasets or hybridized architecture where quantum processing is presented as auxiliary that it can be implemented in fraud prevention. The classical method is the practical one until quantum hardware and the quantum kernel algorithm have developed, and QSVC will be a new opportunity in the mapping of high-dimensional low-latency anomaly detection applications in the future. World Journal of Advanced Research and Reviews, 2025, 28(02), 947-961 955 Figure 4 Side-by-side comparison of Random Forest (this study) and QSVC reported in [31] across standard metrics Figure 4 presents a performance comparison of performance between the previous studies on the implementation of Random Forest and QSVC (Quantum Support Vector Classifier). Although Random Forest model does have a perfect performance in terms of all measurements, QSVC as well demonstrates existing promising results, but it still has some current limitations, such as hardware and scalability, which are unable to perform the same level of model performance in large scale applications. 4. Discussion 5. Interpretation of Results Random Forest model outdid itself because of the performance, which is remarkable in terms of the detection rate, and the accuracy is 99.95, F1 score of 0.8256, and an AUC-ROC which is 0.9759 [32]. The above metrics imply that the model is quite resourceful in distinguishing between fraud transactions and non-fraud transactions. The Confusion Matrix also indicated that no false positives or false negatives had been detected by the Confusion Matrix because all the fraudulent transactions had been identified correctly. This outcome is especially significant when it comes to fraud detection, as the reduction of false positives and false negatives is an essential consideration. False positives might give an unwarranted intervention, whereas false negatives might represent someone escaping with a fraud. The balance of metrics of high accuracy and balanced performance demonstrates the reliability and robustness of the model to work with in practice. Also, the AUC-ROC value 0.9759 shows that the model captures a high true positive rate with a low false positive rate, which further establishes the effectiveness of the model in separating the two classes. As it was determined in the course of the analysis of feature significance, the most powerful predictors of fraudulent transactions were the number of transactions, transaction time interval, and place. Continually, the model rated them among the top 15, which demonstrates the importance of these features in detecting fraudulent practices. These are the major characteristics that give insights that are beneficial to financial institutions. As an example, an unusual size of transaction, a back-to-back transaction, or a transaction made in unusual places can be effectively used as a tool to determine fraud. Such information will help to take proactive action to prevent fraud and increase the protection level of financial systems. The Synthetic Minority Over-sampling Technique (SMOTE) was used to manage the existence of the class imbalance that existed in the fraud detection data. SMOTE is used to create artificial samples of the minority group, thus balancing the data and enhancing the model to learn between the two classes [33]. The use of SMOTE made the dataset more balanced, which resulted in better model performance. It is worth noting, though, that in some cases the use of synthetic data will result in overfitting, in which the pattern that the model has