Researchers Discover Automated Circuit Screening Achieves High Performance for Fraud Detection with 97% Accuracy

Detecting credit card fraud presents a persistent challenge, demanding increasingly sophisticated methods to counter evolving attack patterns and imbalanced datasets, and researchers are now exploring the potential of quantum machine learning to address this need. Nouhaila Innan from New York University Abu Dhabi, Akshat Singh from New York University Tandon School of Engineering, and Muhammad Shafique, also from NYUAD, introduce CircuitHunt, a novel framework that automates the search for effective quantum circuits for fraud detection. This system efficiently screens a vast library of circuits, rapidly identifying those best suited to the task and significantly reducing the time required to find high-performing models from days to hours. The team demonstrates that CircuitHunt achieves 97% test accuracy and a strong macro-F1 score on a demanding fraud detection benchmark, offering a scalable and practical tool for deploying quantum machine learning in critical financial applications.

This rapidly evolving field investigates adapting standard machine learning models to quantum circuits and developing advanced techniques like automated quantum architecture search and federated learning, demonstrating significant potential for improving accuracy and efficiency. Researchers are implementing and refining QML models, including Quantum Feature Deep Neural Networks, Quantum Support Vector Machines, and Quantum Long Short-Term Memory networks, to analyze financial data. A key focus is on quantum feature engineering, where quantum circuits extract and transform data features to enhance model performance.

Furthermore, the field actively develops methods for automatically designing optimal quantum circuits for fraud detection, utilizing techniques like reinforcement learning and evolutionary algorithms. To address data privacy concerns, researchers are exploring Federated Learning in conjunction with QML, enabling model training on decentralized data sources without sharing sensitive information. They are also investigating other privacy-enhancing technologies, such as differential privacy. Publicly available datasets, including the Credit Card Fraud Detection dataset and the KetGPT dataset, are used for training and evaluating these models. A central goal is to develop resource-efficient QML models suitable for near-term quantum devices, minimizing the number of qubits and quantum gates required for practical implementation.

Automated Quantum Model Discovery for Fraud Detection

Researchers have developed CircuitHunt, a fully automated framework designed to discover high-performing quantum machine learning (QML) models, addressing the challenges of scalability and efficiency in quantum architecture search. The team carefully prepares imbalanced fraud detection datasets, using techniques like Synthetic Minority Oversampling to increase the representation of minority classes, and partitions data into training, validation, and test sets using stratified sampling to maintain class proportions. CircuitHunt filters circuits from the KetGPT dataset based on hardware constraints and architectural properties, focusing on circuits containing between 3 and 10 qubits with at least one trainable gate and a limited number of trainable parameters. Candidate circuits are embedded into a standardized hybrid quantum neural network and evaluated through brief training runs, allowing for rapid assessment and elimination of underperforming architectures. The top-ranked circuits are fully trained, achieving 97% test accuracy and a high macro-F1 score on a challenging fraud detection benchmark. This approach significantly reduces architecture search time, providing a scalable tool for QML deployment in financial applications.

CircuitHunt Automates Viable Quantum Machine Learning Models

Researchers developed CircuitHunt, a fully automated framework that streamlines the discovery of high-performing quantum machine learning (QML) models, particularly for credit card fraud detection. The team found that manually designed or randomly initialized quantum circuits often fail to produce meaningful gradients or converge to acceptable performance, while a large majority of circuits generated by KetGPT were trainable and executable, demonstrating a substantial improvement in initial viability. CircuitHunt employs a constraint-aware filtering pipeline that excludes circuits exceeding qubit or parameter limits, ensuring only viable candidates are considered for evaluation. The framework embeds each circuit within a standardized hybrid QNN and performs brief training runs, efficiently eliminating underperforming architectures based on macro-F1 score validation.

This approach delivers 97% test accuracy on a fraud detection benchmark, while maintaining a high macro-F1 score. Ablation studies confirm the contributions of skip connections and constraint-based filtering to the overall success of CircuitHunt. The team emphasizes the need for scalable, automated methods as datasets diversify and application demands become more rigorous.

CircuitHunt presents a novel, automated framework for discovering effective quantum circuits, specifically designed for challenging tasks like credit card fraud detection. The system efficiently filters circuits from a large dataset, rapidly evaluates their performance using macro-F1 scores, and ultimately identifies high-performing models that achieve strong accuracy and balanced performance on fraud detection benchmarks. This approach significantly reduces the time required for architecture search, while maintaining competitive results. The research demonstrates the potential of data-driven quantum circuit discovery, offering a scalable alternative to manual design and providing valuable architectural insights. However, the authors acknowledge limitations related to computational scalability and the need to assess performance in real-world hardware environments, as the current evaluation relies on simulations.