Comparative Performance Analysis of Quantum Machine Learning Architectures for Credit Card Fraud Detection

Mansour El Alami, Nouhaila Innan, Muhammad Shafique, Mohamed BennaiarXiv 2024 · DOI · arXiv

Abstract

Financial institutions face significant challenges in detecting credit card fraud due to high dimensionality and extreme class imbalance in transactional data. This study provides a rigorous comparative analysis of three hybrid Quantum Machine Learning (QML) architectures—Variational Quantum Classifiers (VQC), Sampler Quantum Neural Networks (SQNN), and Estimator Quantum Neural Networks (EQNN). By evaluating these models across various feature maps and ansatze on real-world and synthetic datasets, we demonstrate that VQC and SQNN can achieve F1-scores up to 0.88. Our findings highlight that the choice of data encoding and circuit structure is the primary determinant of performance, with entanglement-based feature maps providing a distinct advantage in capturing complex fraud patterns.

Introduction

Financial fraud represents a multi-billion dollar global threat. While classical machine learning (ML) has improved detection rates, it often struggles with the evolving patterns and high-dimensional nature of modern financial data. Quantum Machine Learning (QML) offers a promising alternative, leveraging quantum entanglement and superposition to process complex data distributions more effectively. However, the design space for QML is vast, and identifying which specific configurations are suitable for Noisy Intermediate-Scale Quantum (NISQ) hardware remains a critical challenge. This research systematically benchmarks different QML architectures to determine their practical utility in financial security.

Experimental Methodology

The study evaluates three distinct hybrid quantum-classical architectures using a comprehensive experimental pipeline. The workflow involves data preprocessing, feature reduction via Principal Component Analysis (PCA), and a systematic exploration of quantum circuit components.

Figure 1: Overview of the experimental workflow, including data preprocessing, model components, and performance evaluation metrics.

Datasets and Preprocessing

Two datasets were utilized to ensure the robustness of the findings:

  • BankSim: A synthetic dataset based on real-world bank payments in Spain, comprising nearly 600,000 records.
  • European Credit Card Dataset: A real-world dataset containing 284,807 transactions from European cardholders.

To address the inherent class imbalance in fraud data, both datasets were balanced using Random Under Sampling (50/50 ratio) and scaled to a [0, 1] range to meet the requirements of quantum gate operations.

Quantum Circuit Design

The performance of the QML models was tested against a matrix of configurations:

  1. Feature Maps (Data Encoding): Z, ZZ, and Pauli maps were used to encode classical data into quantum states, with ZZ and Pauli maps incorporating varying levels of entanglement.
  2. Ansatz (Trainable Layers): Four specific circuit structures were evaluated for their ability to represent the target function: Real Amplitudes, Efficient SU2, Two Local, and Pauli Two Design.
Figure 5: Visual representation of the four considered ansatze (trainable quantum layers) used in the study.

Comparative Analysis of Architectures

Variational Quantum Classifier (VQC)

The VQC utilizes a quantum circuit to map data and a classical optimizer (COBYLA) to tune parameters. It emerged as the top performer, achieving an F1-score of 0.88 on the European dataset when paired with a Z-feature map and Pauli Two Design ansatz. Notably, the VQC demonstrated superior stability under realistic quantum noise conditions.

Sampler Quantum Neural Network (SQNN)

The SQNN utilizes a quantum sampler to extract bitstring probabilities. It was the most consistent performer across both datasets, maintaining a median F1-score near 0.80. Its consistency suggests it may be more resilient to variations in data distribution than other models.

Estimator Quantum Neural Network (EQNN)

The EQNN feeds quantum expectation values into a classical fully connected neural network. Surprisingly, this model performed the worst. The researchers attribute this to the difficulty of mapping expectation values to discrete class labels and limited circuit expressivity in this specific hybrid configuration.

Key Insights and Discussion

  • The Role of Entanglement: Feature maps incorporating entanglement (ZZ and Pauli) consistently outperformed the simpler Z-feature map. This suggests that the "quantum advantage" in fraud detection stems from the ability to capture complex, non-linear correlations that classical models might overlook.
  • Statistical Significance: Analysis of Variance (ANOVA) revealed that the Model Type was the most significant factor in performance ($p < 10^{-102}$), whereas the specific dataset used was not statistically significant. This implies that a well-designed QML architecture is robust across different types of financial data.
  • Noise Resilience: In robustness tests, the VQC model occasionally showed a performance peak under certain levels of phase-flip noise, suggesting a potential regularization-like effect where a small amount of noise prevents overfitting.

Conclusion

This research demonstrates that QML architectures, particularly VQC and SQNN, are viable candidates for credit card fraud detection in the NISQ era. The study underscores that architectural design—specifically the integration of entanglement in feature maps—is more critical than the specific dataset used. Future work will focus on implementing error mitigation techniques like Zero-Noise Extrapolation (ZNE) and exploring deeper circuits to further enhance detection accuracy on physical quantum hardware.

Summary updated

BibTeX
@misc{alami2024comparative,
  author = {Mansour El Alami and Nouhaila Innan and Muhammad Shafique and Mohamed Bennai},
  title = {{Comparative Performance Analysis of Quantum Machine Learning Architectures for Credit Card Fraud Detection}},
  year = {2024},
  eprint = {2412.19441},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.19441},
}
Figure