The application area: when detection drowns in false positives
Every regulated financial institution must run automated fraud and AML/CFT detection, and everywhere in the world, those systems fail in the same way. The models are trained on narrow, historic data; the alerts they generate are overwhelmingly false; and the burden lands on compliance teams and ordinary customers alike.
- The alert crisis. False-positive rates in transaction monitoring run as high as 99%, and alert review can consume up to 40% of a compliance team's time, cost without capture.
- The imported-model problem. Institutions buy detection models trained on other countries' transaction patterns. Models that have never seen an economy's real behavior flag ordinary life as suspicion.
- The inclusion cost. The response to noise is defensive de-risking: accounts closed, correspondent relationships cut. Research published by our co-founder documents how AML/CFT algorithms disproportionately exclude disadvantaged groups (Grillaert, IEEE ETHICS 2025).
- The data deadlock. Detection would improve if institutions could learn from each other, but privacy law across jurisdictions (GDPR, HIPAA and national data protection acts) rightly forbids pooling customer data. So every institution trains alone, on a partial view of the threat.
This is the application area: detection is mandatory, collaboration is impossible, and the gap is filled with false positives. Federated learning resolves it, institutions train a shared model while customer data never moves.
The application: KYComm
KYComm, Know Your Community, is Goose FL's application of the finance use case to AML/CFT fraud detection and false-positive reduction. A shared detection model is sent to every participating institution. Each trains it behind its own walls, on its own transactions and fraud labels. Only encrypted, privacy-protected mathematical updates return; they are combined on sovereign ground, and the improved model flows back, personalized to each institution's customer base.
- Detection that reads local reality. Personalized models (pFedMe) learn the informal-sector patterns imported models misread, the market trader cycling cash to restock her stall is commerce, not crime. False positives are cut where they start.
- Privacy that holds at every layer. Differential privacy under a formal, regulator-visible budget (ε ≤ 8.0 cumulative, materially stronger than deployed practice such as Apple's ε = 4 or the US Census' 15.34); secure aggregation so no update can be traced to any customer; no central vault of anyone's data.
- Oversight without raw data. A real-time supervisory (SupTech) dashboard gives regulators model performance, privacy-budget consumption and alert quality, never a customer record. Every training round is logged to a tamper-evident audit trail.
- Data rights by design. Erasure requests are executed inside the trained model, certified unlearning or retraining, with a completion certificate per request.
- Strengthens, never replaces. The institution's compliance team reviews every alert; CDD, sanctions screening and STR filings to the Nigeria Financial Intelligence Unit remain fully with the institution.
The evidence, by the numbers
The platform was validated using authentic transaction datasets. Benchmark testing on the ULB European credit card fraud dataset demonstrated its performance across 284,807 real-world transactions.
Federated machinery validated on the ULB benchmark (284,807 real transactions) and the IEEE-CIS dataset with three simulated banks and non-IID partitions. Full methodology, results tables and the synthetic-Nigeria validation pipeline are in the technical white paper below.
What this means for institutions
Banks and financial institutions gain automated detection that meets the mandate, with fewer false positives, defensible audit trails, and the biggest lift where local data is thinnest. KYComm is built to support compliance with the CBN Baseline Standards for Automated AML Solutions, including its requirements for model governance and the explainability of automated detection decisions, with coverage extending across further controls in the standards. Regulators and the Nigeria Financial Intelligence Unit gain real-time sector visibility without raw-data access, and higher-quality STRs upstream instead of more noise. Customers keep their access to the formal economy.
Watch: the Compliance Paradox, on film
Our session with the Center for Financial Inclusion, recorded in Abuja with regulators, policymakers and federated-learning researchers, examines the Compliance Paradox and the KYComm response.
The technical white paper
The Compliance Paradox: How Federated Learning Resolves the AML/CFT Deadlock
Building detection that includes your customers?
Whether you're a bank, a fintech or a regulator, we co-design and build this with institutions.
Talk to Goose FL