Research program · Fraud and risk AI
Credit unions and community banks face the same synthetic identity and bust-out fraud as the largest banks, with a fraction of the data to fight it. I am building a way for small lenders to detect fraud together without any member data leaving their own systems.
In active research · Now inviting pilot partners
The problem
A lender with a few hundred thousand members may see only a handful of confirmed synthetic identity cases a month. That is too few to train an accurate model, so most small institutions rely on rules and manual review that miss new schemes and slow down good applicants.
Pooling data with peers would fix the volume problem, but privacy law and member trust make sharing raw records a non-starter. Organized rings exploit the gap by moving from one small institution to the next, knowing no single lender can see the pattern.
The approach
Each institution trains a fraud model on its own systems. Only encrypted model updates leave the building, and they are combined into a shared model that every participant benefits from. No member records are pooled, copied or exposed.
Standard federated learning breaks down on fraud, where confirmed cases are extremely rare and every lender's members and data look different. The research is focused on making it work under exactly those conditions.
Research focus
Each one is an open technical question. Together they decide whether small lenders can match the fraud detection of the largest institutions.
Getting a shared model to catch fraud that makes up well under one percent of applications, without drifting toward the largest participants and missing patterns at smaller ones.
Class imbalance · Personalized models · Fraud capture at fixed false-positive ratesFlagging when the same synthetic identity, or a near variant of it, applies at several institutions at once, without revealing who the applicant is.
Privacy-preserving record linkage · Fuzzy matching · Synthetic identityMeasuring how much detection survives formal privacy protections, and directing that protection where it costs the least accuracy on rare cases.
Differential privacy · Privacy budgets · Detection tradeoffsLimiting the influence any single institution can have on the shared model, so poor or manipulated updates can't corrupt it.
Robust aggregation · Poisoning resistance · Participant qualityWho it's for
Background
This program grew out of a pattern I saw again and again: the institutions most exposed to fraud are the ones with the least data to fight it.
I have worked hands-on in fraud and credit-risk analytics at American Express, Citi, Vanguard and Teachers Federal Credit Union, covering first-party and bust-out fraud, synthetic identity, account takeover and application fraud, and the tradeoff between fraud capture and false positives. At American Express, I built a global data science team from zero to 32 people across four markets.
Pilot partners
I am looking for a small group of credit unions, community banks and industry partners to advise on the research and take part in an early pilot. Participation starts with a conversation, and no member data is shared at any stage.