Q-SYNTH: Hybrid Quantum-Classical Adversarial Augmentation for Imbalanced Fraud Detection

This paper introduces Q-SYNTH, a hybrid quantum-classical generative adversarial framework that utilizes a parameterized quantum circuit to synthesize fraudulent transaction samples, effectively addressing class imbalance in credit card fraud detection by offering a favorable compromise between statistical distributional fidelity and downstream classification performance compared to classical baselines.

The Gist

The Problem: The "Needle in a Haystack"

Imagine you are a security guard at a massive airport. Your job is to spot terrorists (fraudsters) among millions of regular travelers (legitimate customers).

• The Reality: For every 10,000 people walking through, maybe only 5 are actually trying to do something bad.
• The Mistake: If you train a computer to spot these bad guys using only the real data, the computer gets lazy. It learns that "everyone is good," so it just guesses "Good" for everyone. It gets a 99.9% score on its test, but it misses every single bad guy. This is called class imbalance.

The Old Solutions: "Copy-Paste" vs. "Fake It"

To fix this, experts try to give the computer more examples of bad guys.

1. SMOTE (The "Copy-Paste" Method): Imagine taking a photo of a bad guy and drawing a straight line to another bad guy, then creating a new photo right in the middle. It's safe and looks very similar to the real thing, but it's a bit boring and doesn't show the full variety of how bad guys might act.
2. Classical GANs (The "Art Forger"): This uses a computer program that tries to "forge" fake bad guy profiles. One part of the AI (the Generator) tries to make a fake ID, and another part (the Discriminator) tries to catch the fake. They play a game of cat-and-mouse. While this creates very diverse fakes, sometimes the forgeries are a little too obvious or don't match the real statistics perfectly.

The New Solution: Q-SYNTH (The "Quantum Art Forger")

This paper introduces Q-SYNTH, a new hybrid system. Think of it as a team-up between a human artist and a quantum robot.

• The Generator (The Quantum Artist): Instead of using a standard computer brain, this part uses a Quantum Circuit. Imagine a quantum computer as a magical paintbrush that can mix colors in ways a normal brush can't. It creates new, fake fraud profiles that are mathematically complex and diverse.
• The Discriminator (The Human Art Critic): This part is a standard, classical computer (like the ones we use today). Its job is to look at the real fraud profiles and the quantum-generated fake ones and try to tell them apart.

They play a game: The Quantum Artist tries to make fakes so good the Human Critic can't tell them apart. The Human Critic tries to get better at spotting the fakes. Over time, the Quantum Artist gets incredibly good at creating realistic fraud patterns.

How They Tested It

The researchers didn't just say "it works." They ran a strict test with three specific goals:

1. Does it look real? (Statistical Fidelity): They checked if the fake data matched the real data's "shape" (like checking if the fake ID photos have the same distribution of eye colors and heights as the real ones).
   • Result: The Quantum Artist (Q-SYNTH) created fakes that were much closer to the real data than the standard "Art Forger" (Classical GAN), though the "Copy-Paste" method (SMOTE) was still the closest in simple statistics.
2. Can a robot tell them apart? (Detectability): They trained a separate robot to try to spot which data was real and which was fake.
   • Result: The robot was basically guessing (50/50). This is good! It means the fake data is so realistic that even a computer can't easily distinguish it from the real thing.
3. Does it help catch fraud? (Downstream Performance): They used the fake data to train a fraud detector and saw if it caught more bad guys.
   • Result: The Quantum Artist's data helped the detector catch more fraud than the "Copy-Paste" method. While the standard "Art Forger" (Classical GAN) was sometimes slightly better at catching fraud, the Quantum Artist offered a great balance: it was much better at looking like real data and still very good at helping catch fraud.

The "Volume Knob" Experiment

The researchers also tested how much fake data to add. They found that adding a little bit of fake data didn't help much. But when they added a moderate to high amount (about 50% fake, 50% real), the fraud detector got significantly better at its job.

The Bottom Line

Q-SYNTH is a new tool that uses quantum computing to create "fake" fraud data that is incredibly realistic.

• It fixes the problem where computers ignore rare fraud cases.
• It creates data that is statistically very close to the real thing (better than standard AI methods).
• It helps fraud detectors catch more bad guys without needing more real-world data.

The paper concludes that this "Hybrid" approach (Quantum Generator + Classical Critic) is a promising middle ground: it offers the statistical accuracy of simple methods and the powerful learning capability of complex AI, making it a strong candidate for fighting financial fraud.

Technical Summary

Technical Summary: Q-SYNTH

Problem Statement
Credit card fraud detection is fundamentally hindered by extreme class imbalance, where fraudulent transactions constitute a tiny fraction of data (often <1%). Standard supervised learners prioritize the majority (legitimate) class, achieving high overall accuracy but suffering from poor fraud-class recall and F1-scores. While synthetic data augmentation (e.g., SMOTE, GANs) is a common remedy, existing methods face a tension between distributional fidelity (how well synthetic samples match real statistical properties) and downstream performance (how well they improve fraud detection). Classical interpolation methods like SMOTE preserve local statistics but lack diversity, while classical Generative Adversarial Networks (GANs) offer diversity but often struggle with marginal distribution matching and stability in highly imbalanced tabular settings. Furthermore, existing literature often prioritizes downstream metrics without rigorously auditing the statistical similarity between real and synthetic fraud distributions.

Methodology: Q-SYNTH
The paper proposes Q-SYNTH, a hybrid classical–quantum adversarial framework designed to synthesize minority-class fraud samples in tabular data. The architecture consists of:

1. Preprocessing: Raw transaction data is transformed into a bounded, low-dimensional representation space. This involves feature selection (top 10 features), standardization, Principal Component Analysis (PCA) to reduce dimensions to 4, and normalization to the range $[-1, 1]$.
2. Generator (Quantum): A parameterized quantum circuit (PQC) acts as the generator. It takes a latent vector, maps it via a classical neural network to circuit parameters, and prepares a quantum state using angle embedding (Y-rotations) followed by variational layers containing single-qubit rotations ($R_X, R_Y, R_Z$) and ring entanglement (CNOT gates). The output is obtained via Pauli-Z expectation values, naturally producing bounded outputs consistent with the preprocessed domain.
3. Discriminator (Classical): A classical neural network distinguishes between real and generated samples.
4. Training Protocol: The framework employs a stabilized adversarial training loop featuring:
   • Instance Noise: Injected into both real and generated samples to prevent discriminator overfitting.
   • Regularization: Combines adversarial loss with feature matching (aligning intermediate discriminator features) and moment matching (aligning batch-wise means and standard deviations).
   • Adaptive Tuning: Regularization parameters (noise scale, label smoothing, dropout) are dynamically adjusted based on discriminator behavior to maintain informative gradients.

Key Contributions

• Hybrid Framework: Introduction of Q-SYNTH, where a variational quantum circuit generates minority-class samples guided by a classical discriminator.
• End-to-End Pipeline: A controlled workflow from preprocessing and bounded representation mapping to inverse transformation for downstream evaluation.
• Unified Evaluation Protocol: A rigorous assessment framework that jointly evaluates distributional fidelity (using Kolmogorov–Smirnov statistics, Wasserstein distances, and real-vs-synthetic detectability via AUC-ROC) and downstream performance (fraud recall and F1-score) across both quantum and classical classifiers.
• Empirical Trade-off Analysis: Demonstration of a specific compromise where the hybrid model improves statistical fidelity over classical GANs while maintaining competitive detection performance.

Results

• Distributional Fidelity: In terms of marginal similarity to real fraud data, SMOTE achieved the highest fidelity (lowest KS and Wasserstein distances) due to its interpolation nature. However, Q-SYNTH significantly outperformed the classical GAN baseline, reducing the marginal distribution gap (e.g., Q-SYNTH KS median: 0.069 vs. Classical GAN: 0.185). Crucially, Q-SYNTH maintained low real-vs-synthetic detectability (AUC $\approx$ 0.475, close to random chance), indicating the synthetic samples are not easily distinguishable from real ones by an external classifier.
• Downstream Performance: When used to augment training data, Q-SYNTH improved fraud recall and F1-scores across Quantum Neural Networks (QNNs) and classical classifiers (ANN, Logistic Regression, Random Forest, XGBoost) compared to imbalanced baselines and often compared to SMOTE. While the classical GAN occasionally achieved the highest absolute downstream scores in specific configurations, Q-SYNTH remained highly competitive.
• Scaling Analysis: Performance gains were non-monotonic at low injection ratios (10% synthetic data sometimes degraded performance). Optimal results were observed at moderate-to-high injection ratios (50% and 100%), suggesting that a sufficient volume of synthetic data is required to shift the decision boundary effectively.

Significance and Claims
The paper claims that Q-SYNTH offers a favorable trade-off between statistical fidelity and downstream utility. While classical GANs may maximize downstream performance in some settings, they often suffer from poorer marginal distribution matching. Conversely, Q-SYNTH mitigates the marginal artifacts common in classical adversarial training on tabular data, providing samples that are statistically more faithful to real fraud distributions while still driving significant improvements in fraud detection metrics. The study positions hybrid quantum augmentation as a feasible and promising approach for addressing the specific bottlenecks of imbalanced fraud detection, particularly in low-dimensional tabular representations. The authors note that these results are obtained within a compressed representation pipeline and call for future work on dependence-aware fidelity metrics and hardware-aware evaluation under realistic NISQ (Noisy Intermediate-Scale Quantum) constraints.