Fraud detection in credit card transactions using Quantum-Assisted Restricted Boltzmann Machines

TL;DR

This study assesses quantum-assisted Restricted Boltzmann Machines for fraud detection on Stone's real credit-card dataset, comparing classical training with simulated and actual quantum annealing. By mapping RBMs to QUBO and Ising formulations and integrating quantum sampling, the authors demonstrate potential performance gains on NISQ devices despite noise and embedding constraints. Hyperparameter tuning and dataset balancing are crucial, with quantum approaches achieving competitive metrics at the cost of longer runtimes. The work argues for the viability of quantum-assisted learning in financial anomaly detection and outlines directions for future quantum-accelerated ML methods.

Abstract

Use cases for emerging quantum computing platforms become economically relevant as the efficiency of processing and availability of quantum computers increase. We assess the performance of Restricted Boltzmann Machines (RBM) assisted by quantum computing, running on real quantum hardware and simulators, using a real dataset containing 145 million transactions provided by Stone, a leading Brazilian fintech, for credit card fraud detection. The results suggest that the quantum-assisted RBM method is able to achieve superior performance in most figures of merit in comparison to classical approaches, even using current noisy quantum annealers. Our study paves the way for implementing quantum-assisted RBMs for general fault detection in financial systems.

Figures (3)

• Figure 1: Representation of the probability distribution p(v) before (top) and after (bottom) training the RBM. Probability distribution of the dataset is shown in yellow, while the distribution of the data generated by the model can be seen in blue.
• Figure 2: Representation of a 3x3 Chimera grid, a simplified model of the D-Wave 2000Q quantum annealer. The loops represent the qubits, the blue dots represent the external couplings (which allow for the conection of two unit cells), and the internal couplings are represented by green circles (allowing the conection of one or more qubits inside the unitary cell). Image source: D-Wave's Documentation
• Figure 3: Comparison between the three training methods: strictly-classical PCD (blue), simulated-annealing-assisted training (yellow), and quantum-annealing-assisted training (Purple). Training was executed in 50 epochs; for each epoch, Accuracy, Precision, Recall, and F1-Score were calculated as comparative metrics. Due to limited access to D-Wave's system, only certain values were available in the quantum-annealing-assisted training; they are plotted as purple dots in the figure. The optimum configuration achieved by each training method corresponds to epoch 25, 34, and 49, respectively.