In a paper published in the journal Entropy, researchers developed a fraud detection method using quantum computing to detect communities in transaction networks. Transaction data was modeled as an undirected graph, and the quadratic unconstrained binary optimization (QUBO) model optimized the modularity function to identify communities.
Study: Quantum Computing in Community Detection for Anti-Fraud Applications. Image Credit: Funtap/Shutterstock.com
The coherent ising machine (CIM) outperformed classical methods in speed and modularity quality, successfully placing a high-risk community containing fraudulent accounts. This approach demonstrated significant potential for enhancing anti-fraud strategies in banking.
Related Work
Past work explored quantum computing applications for financial fraud detection, emphasizing its potential in tackling NP-hard optimization problems. Community detection was framed as a QUBO problem to maximize modularity, leveraging the equivalence between QUBO and ising models.
CIMs, capable of solving large-scale optimization at room temperature, were employed to identify fraudulent account communities.
Fraud Detection and Analysis
This study's dataset originates from a Chinese commercial bank's fraud detection scenario. The data collection process begins by selecting fraudulent cases from detected accounts.
Subsequently, all one-step connected nodes involved in transactions with these cases form the first-degree association sample set. This process is repeated iteratively to form second-degree and third-degree association sample sets. The final dataset comprises 3,934 samples, of which 186 are labeled as fraudulent.
Given the prevalence of noise in fraudulent transaction data, a denoising process was applied to improve detection efficacy. High-risk accounts were identified using rule-based methods, and low-risk widespread and isolated nodes were removed from the transaction graph.
The denoised dataset consisted of 308 accounts, 19 of which were fraudulent, increasing the fraud probability to 6.17%. While this approach may miss some individual or occasional fraud cases, it effectively targets organized fraud networks, complementing other rule-based or artificial intelligence (AI) methods.
Beijing QBoson quantum technology Co. Ltd. provided the quantum computing setup utilized in the study. It features an optical component, including a pulsed laser, fiber amplifiers, and periodically poled lithium niobate crystals, alongside an electrical component with balanced homodyne detectors, converters, and field-programmable gate arrays. The system operates using degenerate optical parametric oscillation and implements the ising hamiltonian via feedback control signals modulating laser intensity and phase.
The Louvain and simulated annealing (SA) algorithms were run 100 times each to evaluate community detection. The Louvain algorithm results were derived from the median modularity value. Similarly, the SA algorithm was run with a high initial temperature and gradual cooling, with 1000 iterations per temperature step.
CIM Superiority
The CIM employs field-programmable gate arrays (FPGAs) to compute feedback signals, requiring the digitization of laser pulse amplitudes. It introduces a precision limit to the QUBO matrix. Tests with different encoding precisions demonstrated that both configurations successfully identified the optimal community structure, surpassing traditional algorithms like Louvain and SA in performance and efficiency.
The CIM achieved higher modularity values than Louvain and SA. While one precision configuration performed slightly lower than Louvain, the other showed clear superiority due to reduced sensitivity to noise and precision errors. Additionally, the CIM operated significantly faster than Louvain, with SA being the slowest of the three. The success rate of the CIM in achieving optimal modularity was notably higher than that of Louvain and SA, highlighting its effectiveness in producing reliable results.
The community structures identified by the CIM, Louvain, and SA were broadly consistent, with the CIM and Louvain yielding identical outcomes. These structures were instrumental in identifying a high-risk community with a significantly higher probability than fraud in the overall population. This high-risk group contained the most fraudulent accounts, making it a critical target for fraud detection efforts.
Overall, the CIM demonstrated exceptional capability in accelerating fraud detection and improving detection quality. Its ability to pinpoint high-risk communities offers substantial benefits for commercial banks, enabling them to identify and mitigate fraudulent activities efficiently. This enhanced accuracy and speed position the CIM as a valuable tool for practical applications in fraud detection scenarios.
Conclusion
To sum up, this study addressed the hardware challenges of applying quantum computing to fraud detection by utilizing the CIM for community detection, framed as a QUBO model. The CIM outperformed the Louvain and SA algorithms by achieving faster detection and a higher likelihood of producing optimal solutions.
The results provided significant business value by identifying high-risk communities, covering nearly 70% of fraudulent accounts.
This study demonstrated the potential of quantum computing in solving complex financial problems and extended the application of the Ising machine beyond traditional AI-based methods. These findings suggested promising future applications of quantum computing for fraud detection.
Journal Reference
Wang, Y., et al. (2024). Quantum Computing in Community Detection for Anti-Fraud Applications. Entropy, 26:12, 1026. DOI: 10.3390/e26121026, https://www.mdpi.com/1099-4300/26/12/1026
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.