Hertha项目：识别实时零售支付系统中的金融犯罪模式2025

Project Hertha Identifying financial crimepatterns in real-time retailpayment systems Contents Executive summary Motivation Synthetic data set Combatting financial crime is essentialto maintaining trust in the financialsystem. It has been estimated1that$3 trillion of money laundering andterrorist financing flow through theglobal financial system every year.Addressing this is increasingly urgentasnew technologies are also enablingnew financial crime threats. The experiments were conductedusing acomplex simulated synthetictransaction data set,developed aspart of the project. It includes dataon 1.8 million bank accounts and 308million transactions. The data set wasbuilt using an AI model trained tosimulate realistic transaction patterns.While no real customer data was used inthe exercise, the data set was designedto be representative of an ecosystem ofretail payments in a single jurisdiction. Project Hertha is a joint project betweenthe BIS Innovation Hub’s London Centre andthe Bank of England. The project exploredhow transaction analytics could help identifyfinancial crime patterns in real-time retailpayment systems, while using the minimumset of data points. To evade detection,criminals operatein complex networkswhich includemany accounts across multiple financialinstitutions. Earlier initiatives, includingthe BIS Innovation Hub’s Project Aurora,demonstrated the potential of networkanalytics to identify this activity innetwork-wide data. Electronic payment systems processtransactions across many participants,which gives them a network-wide view.Project Hertha testedthe applicationof modern artificial intelligence (AI)techniques to help spot complex andcoordinated criminal activity inpayment system data.It measuredthe added value of such transactionanalytics relative to a modelledbenchmark of banks and paymentservice providers (PSPs) monitoringaccounts in isolation. Executive summary The results demonstrate promise,but also show there are limits to theapplication and effectiveness of systemanalytics. It is justone piece of thepuzzle.The introduction of a similarsolution would also raise complexpractical, legal and regulatory issues.Analysing these was beyond the scopeof Project Hertha. The concept exploredin the project does not assume anychanges in the responsibilities ofindividual institutions. Findings –Likewise, the ongoing effectiveoperation of payment systemanalytics requiresbanks and PSPsto continuously provide feedbackon outcomesfor accounts flaggedby the model. Project Hertha found that paymentsystem analytics could be a valuablesupplementary tool to help banksand PSPs spot suspicious activity.Key findings from the project include: The results have beenachieved while using aminimal number of datapoints, demonstrating thatadvanced models can drawon network patterns ratherthan personal data. –Explainable AI approachescould provide additional valuableinformationto aid banks and PSPsin investigations and reporting,such as reasons why an accountwas flagged. –Working in isolation,paymentsystem operators identified fewerillicit accountsrelative to banks andPSPs (39% vs 44%). –Using findings from payment systemanalyticshelped banks and PSPsfind 12% more illicit accountsthanthey would otherwise have found. Key insights Further experiments could test similarapproaches for cross-border andlarge-value payment systems as wellas cryptoasset networks. These wereout of scope for Project Hertha. Results have also pointed at a fewhelpful practical insights: –Payment system analytics wasparticularly valuable for spottingnovel financial crime patterns.When trying to spot previouslyunseen behaviours, it helpedachieve a 26% improvement. –Payment system analytics provedmost effective when targeted atidentifying more complex schemesinvolving many accounts acrossdifferent banks and PSPs. For someschemes, it doubled detection accuracy. The results have been achieved whileusing a minimal number of data points,demonstrating thatadvanced modelscan effectively draw on networkpatterns rather than personal data.They also assume that no private dataare shared with the payment systemoperator. –To achieve the best results,algorithms need to be trained onconfirmed past cases.Unsupervisedalgorithms were found to be farless effective. Section 1Motivation and hypotheses Collaboration between financialinstitutions is essential to combatthe rise of financial crime. ProjectHertha focused on the role ofelectronic payment systems. Motivation and hypotheses:Background New technologies provide improvedopportunities to identify and preventfinancial crime, while balancing thecompeting objectives of protectingprivacy and managing operationalcosts. Advanced artificial intelligence(AI)-enabled models can help toidentify complex patterns in the data.Meanwhile, synthetic data generationcan enable training models moreeffectively, particularly where there arelegal or practical barriers to obtainingreal data. Combatting fin