QFNN-FFD: Quantum Federated Neural Network for Financial Fraud Detection

TL;DR

The paper addresses the need for privacy-preserving, high-accuracy financial fraud detection by integrating quantum machine learning with federated learning into the QFNN-FFD framework. It employs a quantum neural network circuit with angle encoding and a CNOT entanglement scheme, optimized via Adam and parameter-shift gradients, and trained through federated averaging across IID clients. Empirical results on the IEEE-CIS Fraud Detection dataset show precision above 95% and robust performance under multiple quantum noise models, outperforming existing QML fraud detectors while preserving data privacy. The work demonstrates the practicality of privacy-aware quantum-enhanced fraud detection and suggests extensions to other privacy-critical domains such as healthcare and cybersecurity.

Abstract

This study introduces the Quantum Federated Neural Network for Financial Fraud Detection (QFNN-FFD), a cutting-edge framework merging Quantum Machine Learning (QML) and quantum computing with Federated Learning (FL) for financial fraud detection. Using quantum technologies' computational power and the robust data privacy protections offered by FL, QFNN-FFD emerges as a secure and efficient method for identifying fraudulent transactions within the financial sector. Implementing a dual-phase training model across distributed clients enhances data integrity and enables superior performance metrics, achieving precision rates consistently above 95%. Additionally, QFNN-FFD demonstrates exceptional resilience by maintaining an impressive 80% accuracy, highlighting its robustness and readiness for real-world applications. This combination of high performance, security, and robustness against noise positions QFNN-FFD as a transformative advancement in financial technology solutions and establishes it as a new benchmark for privacy-focused fraud detection systems. This framework facilitates the broader adoption of secure, quantum-enhanced financial services and inspires future innovations that could use QML to tackle complex challenges in other areas requiring high confidentiality and accuracy.

Figures (8)

• Figure 1: Comparison of ML and FL accuracies in classical and QC contexts across various fields and experiments. Panel (a) illustrates the performance of different experiments within the finance sector. Panel (b) compares QML with QFL across four domains: healthcare, IoT, computer vision, and finance. In classical computing contexts, FL generally demonstrates superior performance compared to ML shingi2020federatedwang2024novel. In QC contexts, QFL exhibits slight improvements over QML qu2023dtqflqu2024qfsmchen2021federated. These findings highlight the potential of QFL and provide a compelling rationale for its adoption, particularly in the finance sector.
• Figure 2: The QFNN-FFD process flow. The diagram outlines the end-to-end workflow from input through to output. Datasets are processed and fed into the QFNN-FFD, built upon the PennyLane library. The model undergoes training and testing for 100 iterations, incorporating a variety of noise models using noise simulators from IBM's Qiskit. The quantum simulator within PennyLane is utilized to emulate a quantum environment. The output is evaluated based on performance metrics, including accuracy, precision, recall, F1 score, and mean squared error loss, providing a comprehensive assessment of the model's capability to detect fraudulent transactions.
• Figure 3: Schematic representation of the FL architecture. The diagram shows multiple users (clients), each with their local dataset, independently training local models. These models are then transmitted as model updates to a central server. The server aggregates these updates to improve the global model, which is then distributed back to the users for further refinement. This cycle ensures data privacy and security, as raw data never leaves the local premises of each user.
• Figure 4: An overview of the QFNN-FFD framework. This flowchart presents the multi-stage process, beginning with data preprocessing and distribution to various users. Each user independently conducts a local training phase on a QNN circuit, followed by an optimization stage. The optimized local models are then transmitted to a central cloud server for global aggregation, culminating in an enhanced federated model. The lower part of the figure illustrates the quantum circuit's structure, showcasing the intricate interplay of qubits and quantum gates (rotations and CNOT gates) during the computation process.
• Figure 5: General schematic of a QML model workflow. The process begins with qubits in the zero state $(|0\rangle)$. The qubits undergo data encoding to represent the input data in quantum states. Then, a parametrized quantum circuit, $U(\theta)$, transforms the qubit states, where $\theta$ represents tunable parameters. The transformed quantum states are measured, converting quantum information into classical output. This output is evaluated using a predefined loss function, and a classical optimization algorithm iteratively adjusts $\theta$ to minimize the loss, thereby refining the QML model's performance.