18/08/2025
๐๐ก๐๐ฅ๐ฅ๐๐ง๐ ๐ ๐ญ๐จ ๐๐ ๐๐ฎ๐๐ง๐ญ๐ฎ๐ฆ ๐๐๐๐๐ฒ โ ๐๐ซ๐จ๐ฃ๐๐๐ญ ๐๐ฉ๐๐๐ญ๐
Organized by I-HUB QTF & NVIDIA
โจ Weโre excited to share an update from one of our featured projects, โ๐๐๐๐ฅ๐๐๐ฅ๐ ๐๐ซ๐๐๐ข๐ญ ๐๐๐ซ๐ ๐๐๐ฒ๐ฆ๐๐ง๐ญ ๐
๐ซ๐๐ฎ๐ ๐๐๐ญ๐๐๐ญ๐ข๐จ๐งโ, led by Syamasundar Gopasana and Swatantra Singh from Accenture.
๐๐ก๐๐ญโ๐ฌ ๐๐๐ฉ๐ฉ๐๐ง๐ข๐ง๐ ?
The team is exploring the convergence of quantum computing, quantum-inspired computing, and GPU-accelerated classical models to detect fraudulent credit card transactions using real-world financial datasets.
Phase 1 leveraged the BankSim dataset with 4-qubit models, while Phase 2 has scaled to the IEEE-CIS Fraud Detection dataset, using 10-qubit systems and more complex quantum-inspired models.
๐๐๐๐ก๐ง๐ข๐๐๐ฅ ๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โผQuantum SVM with Quantum Kernel Estimation (QKE) using Qiskit
โผcuTensorNet-powered quantum-inspired SVMs with MPI-based multi-GPU parallelism
โผClassical models like Random Forest, XGBoost, and GPU-accelerated SVMs (via cuML)
โผAchieved up to 94.7% ROC-AUC with 0% false negatives on 4-qubit models
โผEfficient simulation strategies with CuPy-based kernel computation on Phase 2
๐๐๐ฌ๐ฌ๐จ๐ง๐ฌ & ๐๐๐๐ซ๐ง๐ข๐ง๐ ๐ฌ:
๐นQiskit GPU backends didnโt yield expected performance- simulation remained CPU-bound
๐นTensor network models faced bottlenecks at high feature dimensions
๐นManual optimization and revised matrix ops improved performance
๐นCritical insight: need to track time lost in failed/cancelled GPU jobs for better resource planning.
๐๐ก๐๐ญโ๐ฌ ๐๐๐ฑ๐ญ?
๐ธComplete training and evaluation of models on the IEEE Fraud dataset
๐ธOptimize tensor contraction paths and parallelization further
๐ธDocument findings to demonstrate real-world potential of quantum-inspired fraud detection
This project demonstrates a strong application of ๐ช๐ฎ๐๐ง๐ญ๐ฎ๐ฆ ๐ซ๐๐๐๐ข๐ง๐๐ฌ๐ฌ ๐ข๐ง ๐ญ๐ก๐ ๐๐ข๐ง๐๐ง๐๐ ๐ฌ๐๐๐ญ๐จ๐ซ, pushing boundaries in fraud detection with advanced simulation techniques.
