⚛️ quantum physics

CircuitHunt: Automated Quantum Circuit Screening for Superior Credit-Card Fraud Detection

This paper introduces CircuitHunt, an automated framework that efficiently screens and optimizes quantum circuits from the KetGPT dataset to achieve superior credit-card fraud detection performance with 97% accuracy, significantly reducing architecture search time while addressing challenges like class imbalance.

Original authors: Nouhaila Innan, Akshat Singh, Muhammad Shafique

Published 2026-10-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nouhaila Innan, Akshat Singh, Muhammad Shafique

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of finance, the battle against credit card fraud is a constant, silent war. Every day, millions of transactions flow through banking networks, and hidden within that vast river of data are the rare, dangerous drops of theft. For decades, banks have relied on classical computer programs to spot these anomalies, but the criminals are getting smarter, and the data is becoming too complex for old methods to handle alone. This has led scientists to look toward a new frontier: quantum machine learning. Unlike the computers in our homes, which process information as simple on-or-off switches, quantum computers use the strange laws of physics to hold information in multiple states at once. This allows them to explore many possibilities simultaneously, offering a potential shortcut to solving problems that are too tangled for traditional machines. However, building these quantum programs is notoriously difficult. Designing a quantum circuit—the specific arrangement of instructions that tells the quantum computer what to do—is currently a slow, manual process. Researchers often have to guess which design will work, much like trying to build a complex machine by randomly snapping together gears and hoping it turns.

A team of researchers at New York University Abu Dhabi and New York University Tandon School of Engineering has introduced a new way to solve this design problem, one that removes the guesswork entirely. They created a system called CircuitHunt, an automated tool that hunts for the best possible quantum circuit to detect fraud. Instead of relying on human intuition or random trials, which often lead to failures, CircuitHunt acts as a rigorous filter. It starts with a massive library of over a thousand pre-made quantum circuits, generated by an artificial intelligence trained on real quantum programming code. The system then puts each of these candidates through a series of strict tests. It checks if the circuit is small enough to run on current technology, if it has the right kind of adjustable parts to learn from data, and if it can actually execute without crashing. This process is like a talent show where thousands of contestants audition, but the judges eliminate the vast majority after just a few seconds, keeping only those who show genuine promise.

Once the system has whittled the list down to a handful of viable candidates, it puts them to work on a real-world challenge: identifying fraudulent credit card transactions. The data used for this task is notoriously difficult because fraud is rare; in a typical dataset, fewer than one percent of transactions are actually fake. This imbalance makes it easy for a computer to default to predicting "no fraud" every time, which would be accurate most of the time but useless for catching thieves. To handle this, the researchers trained their automated system to look for a specific kind of success: a score that balances the ability to catch fraudsters without falsely accusing innocent people. The system rapidly trains each candidate circuit for a short period, checking this balance score. If a circuit performs poorly, it is discarded immediately. If it shows promise, it is kept for further testing. This screening process, which would have taken human experts days or weeks to manage manually, was completed by the system in a matter of hours.

The result of this automated hunt was a single quantum circuit that stood out from the rest. When the researchers gave this winning circuit a full training session, it achieved a remarkable level of accuracy. In tests, it correctly identified fraud and non-fraud cases with a 97 percent success rate, a performance that surpassed several other advanced quantum models that had been carefully hand-designed by experts. The study also revealed something important about how these quantum systems learn. The researchers found that the winning circuit included a specific architectural feature—a direct link that allowed information to flow smoothly from the beginning of the process to the end. When they removed this link to see what would happen, the system's performance collapsed, dropping to a much lower accuracy. This proved that the automated discovery process didn't just find a working circuit; it found the right structure to make the quantum computer effective.

The significance of this work goes beyond just catching credit card thieves. It demonstrates that we can move away from the slow, manual design of quantum algorithms toward a future where machines design themselves for specific jobs. By automating the search for the best quantum structures, the researchers have shown that it is possible to find highly effective solutions quickly, even for difficult tasks where data is scarce and unbalanced. While the current tests were run on a simulator rather than a physical quantum machine, the results provide a clear roadmap for how these systems can be deployed in the real world. The study suggests that the future of quantum computing in finance and security will not depend on a few brilliant human designers, but on automated systems that can tirelessly screen thousands of possibilities to find the one that works best.

Technical Summary: CircuitHunt – Automated Quantum Circuit Screening for Credit-Card Fraud Detection

1. Problem Statement

The paper addresses the critical challenge of designing effective Quantum Machine Learning (QML) models for real-world tasks, specifically credit-card fraud detection. This domain presents two primary difficulties:

Extreme Class Imbalance: Fraudulent transactions typically constitute less than 1% of total activity, causing standard supervised learning pipelines to suffer from bias and underperformance.
Architectural Bottlenecks in QML: Unlike classical neural networks, which benefit from established architectures and scalable Neural Architecture Search (NAS), the landscape of Parameterized Quantum Circuits (PQCs) is fragmented. Existing approaches rely on either:
- Manual Design: Hand-crafting circuits based on limited domain knowledge, leading to a lack of structural diversity.
- Random Initialization: Generating random circuits without considering trainability or task suitability, resulting in high failure rates and inefficient resource usage.

The authors argue that current QML research lacks an automated, task-driven framework to efficiently screen and select high-performing quantum circuits from large, structured datasets under practical hardware constraints.

2. Methodology: The CircuitHunt Framework

The authors propose CircuitHunt, a fully automated quantum circuit screening framework designed to discover high-performing models from the KetGPT dataset (a collection of 1,000 structurally valid quantum circuits generated by transformer models). The pipeline operates in three main stages:

A. Data Preprocessing

Imbalance Handling: The credit-card fraud dataset is processed using SMOTE (Synthetic Minority Over-sampling Technique) to rebalance the input distribution.
Normalization: Features are scaled using Min-Max normalization to a bounded interval to ensure compatibility with quantum processing.
Splitting: The data is partitioned into training, validation, and test sets using stratified sampling to preserve class proportions.

B. Quantum Circuit Filtering

The framework filters the initial set of KetGPT circuits ( $C$ ) to create a viable subset ( $C'$ ) based on strict hardware and learning constraints:

Qubit Budget: Circuits must have between 3 and 10 qubits ( $3 \le n \le 10$ ) to align with current simulation and hardware capabilities.
Trainable Gates: Circuits must contain at least one trainable gate (e.g., $RY(\theta)$ , $RZ(\theta)$ , $U2(\phi, \lambda)$ ).
Parameter Budget: The total number of trainable parameters must not exceed 30 ( $P \le 30$ ).
Execution Validation: Candidates are simulated with zero-initialized parameters; only those yielding finite expectation values within bounds are retained.

C. Candidate Evaluation & Selection

Each filtered circuit is embedded into a standardized Hybrid Quantum Neural Network (HQNN) architecture:

Pre-Quantum Classical Network: A feedforward network maps input features to a latent space matching the circuit's qubit width.
Quantum Module: Inputs are encoded via parameterized rotation gates, processed by the candidate KetGPT circuit, and followed by entangling operations (e.g., CNOT chains).
Learnable Skip Connection: A residual connection ( $z_{res} = z_{quantum} + \alpha \cdot z_{classical}$ ) is introduced to preserve gradient flow and enhance expressivity.
Post-Quantum Network: A final MLP outputs logits for binary classification.

Selection Process:

Each candidate is trained for a short duration (5 epochs) using the Adam optimizer and Binary Cross-Entropy loss.
Macro-F1 Score is used as the primary selection metric to ensure balanced performance across both majority (non-fraud) and minority (fraud) classes.
The circuit achieving the highest validation Macro-F1 score is selected, re-trained from scratch for 20 epochs, and evaluated on the test set.

3. Key Contributions

The paper outlines four primary contributions:

Constraint-Aware Circuit Filtering: A pipeline that excludes circuits violating qubit counts, parameter limits, or execution stability, ensuring only viable candidates proceed.
Macro-F1 Score Based Selection: An evaluation strategy that prioritizes balanced class performance over raw accuracy, critical for imbalanced fraud detection.
Modular End-to-End Pipeline: A reproducible workflow integrating preprocessing, quantum module selection (with learnable skip connections), and post-processing.
Ablation and Performance Analysis: Empirical evidence demonstrating that the selected circuit outperforms manual designs and that the residual skip connection is vital for model convergence and generalization.

4. Experimental Results

Experiments were conducted on a MacBook Air (M3 chip) using the PennyLane simulator. The selected circuit (Index #221 from KetGPT) utilized 6 qubits and 9 trainable parameters.

Performance Metrics: The CircuitHunt-selected model achieved:
- 97% Test Accuracy
- 97% Precision, Recall, and F1-Score for both fraud and non-fraud classes.
- ROC-AUC of 0.9946, indicating near-perfect discriminative power.
Comparison: The model outperformed several state-of-the-art manually designed QML architectures (EstimatorQNN, QGNN, SQNN, QFDNN), which showed trade-offs between precision and recall (e.g., SQNN had 100% precision but only 68% recall).
Ablation Study: Removing the residual skip connection resulted in a significant performance drop:
- Accuracy fell to ~70%.
- Macro-F1 score dropped below 0.70.
- ROC-AUC decreased to 0.7706.
- This confirms the skip connection is essential for gradient flow and handling class imbalance.

5. Significance and Claims

The paper claims that CircuitHunt provides a scalable and task-driven tool for QML deployment in critical financial applications. Its significance lies in:

Automation: It reduces architecture search time from days/weeks (manual tuning) to hours by automating the discovery of high-performing circuits.
Robustness: By combining budget-aware pruning with empirical validation, it mitigates the high failure rates associated with random or manual circuit selection.
Practicality: It demonstrates that automated discovery can yield models that are not only structurally valid but also empirically superior to expert-designed counterparts in real-world, imbalanced scenarios.

The authors maintain a modest stance, acknowledging limitations such as the computational cost of the filtering stage and the fact that current evaluations are based on simulated environments rather than noisy real-world quantum hardware. They conclude that while the approach shows strong potential, further investigation is needed regarding transferability to other domains and scalability to larger datasets.