← All projects

A noise-resilient quantum classifier for credit-card fraud, benchmarked against classical models under a realistic shot budget.

Results · full test set, 56,962 transactions, 98 frauds

0.735
AUPRC for NCRC at p = 0.05 noise
0.405
Original VQC at the same noise
1.000
Spearman ρ: ranking preserved
−88%
Hardware executions, frugal schedule
Bar charts of AUPRC and F1 at noise levels 0.01 and 0.05 for Linear, Original VQC, Entangled NCRC and NCRC
AUPRC and F1 at p = 0.01 and 0.05, mean ± std over five seeds.
Precision–recall curves for the four models at two noise levels
Precision–recall curves, seed 0.
01

The problem

Quantum classifiers face noisy hardware and limited measurement budgets. Fraud detection adds an imbalanced dataset, making both prediction quality and computational cost important.

02

What we built

BreaQ develops the Noise-Covariant Re-uploading Classifier (NCRC): four independent quantum wires that preserve transaction rankings under the specified depolarizing-noise model. The pipeline uses PennyLane and PyTorch, compares against classical baselines on the ULB credit-card dataset, and tracks circuit executions and measurement shots.

03

Under the hood

Symmetry-aware circuit design and packed parameter-shift gradients make the classifier resilient to the tested noise model while reducing execution costs. The project also checks finite-shot readout and other noise channels, without claiming quantum advantage over classical machine learning.