Fraud and financial crime

Fraud detection for banks: use cases, data and architecture

Banks detect fraud by scoring every card, transfer and login event in real time against a governed history of customer behaviour, devices and counterparties. The main use cases are card and payment fraud, account takeover, authorized push payment scams, mule accounts, cheque and loan application fraud. A sovereign lakehouse lets one copy of data serve real-time scoring, investigation and model retraining inside the country.

Published . Last updated

On this page

What fraud do banks need to detect?

Most bank fraud falls into six groups. Each needs different data and a different response time.

Fraud typeWhat happensDecision window
Card and payment fraudStolen or cloned card details used online or at a terminalUnder a second, at authorization
Account takeoverA criminal logs in as the customer, often after phishing or SIM swapSeconds, at login or first payment
Authorized push payment scamsThe real customer is tricked into sending moneySeconds, before the transfer releases
Mule accountsAccounts that receive and move stolen moneyHours to days, across the network
Application fraudFake or synthetic identities open accounts or take loansMinutes, at onboarding
Cheque and internal fraudAltered cheques, kiting, staff abuseHours to days

What data does a bank fraud model need?

The model is only as good as the context it sees at decision time. The useful signals are:

  • Transaction history: amount, merchant, channel, time and location, compared with the customer’s own normal pattern.
  • Device and session data: device fingerprint, IP, new device, typing speed, session length.
  • Customer profile: tenure, products held, recent changes to phone, email or address.
  • Counterparties: is the beneficiary new, has it received money from many unrelated customers, is it on an internal watchlist?
  • Telecom signals: a recent SIM swap on the customer’s number is a strong account takeover signal.

Most of these sit in different systems: core banking, card switch, digital banking, CRM. Fraud teams lose most of their time joining them.

How does real-time fraud detection work on a lakehouse?

  1. Connect captures transactions and logins from core and channel systems with change data capture.
  2. StreamHouse streams events into open Delta or Iceberg tables and computes rolling features, such as spend in the last hour or new beneficiaries this week.
  3. Lakebase serves the latest customer features to the scoring service with low latency.
  4. DataNature AI scores each event with a model trained on governed history.
  5. AI360 Studio sends the decision (approve, step-up, hold, decline) back to the payment system and opens a case for analysts.
  6. Atlas records lineage from every alert to the source events, for audit and model review.

The full pattern is in how to build real-time fraud detection on a lakehouse.

Rules or machine learning?

Both. Rules catch known patterns and are easy to explain to auditors. Models catch new patterns and rank alerts so analysts see the riskiest first. Most banks run rules as a safety net and use model scores to decide the action and the queue order.

How do you reduce false positives?

False positives cost money and customers. Three things help most:

  • Compare each event with the customer’s own baseline, not a global average.
  • Use step-up authentication instead of a decline when risk is moderate.
  • Feed analyst outcomes back as labels every week, so the model learns which alerts were real.

Why does sovereignty matter for bank fraud data?

Fraud data is some of the most sensitive data a bank holds: full transaction histories, devices, locations and investigation notes. Central banks and regulators increasingly expect it to stay in the country. With DataNature, data, models and AI prompts stay inside your jurisdiction, on Edwin in Canada, EverestCloud in Nepal, or your own data centre.

Where to start

Start with one high-loss channel, usually card-not-present or instant transfers. Build the features once in the lakehouse, then reuse them for account takeover and mule detection. See the real-time card and payment fraud use case.

How is a bank fraud platform built on DataNature?

  • Hot path: StreamHouse captures card authorizations and transfers from core systems and payment switches, maintains short-window features and feeds real-time scoring.
  • Cold path: Connect loads customer, account and historical transaction tables nightly; dbt builds Silver and Gold, including a feature store of slower behavioural features.
  • Models: trained on Spark, tracked and versioned in the DataNature AI registry, served for real-time and batch inference.
  • Decisions: AI360 Studio combines rules and scores; Voir gives investigators a case assistant scoped to fraud data.
  • Governance: card numbers masked to last four, customer IDs hashed for analysts, investigators exempt; every score and decision written to Iceberg for audit and model risk review.

How does DataNature Atlas keep this secure and compliant?

Investigators see what they need; analysts see masked identifiers; every score and decision is auditable. DataNature Atlas is the platform's governance module:

  • Knows your data: a catalog with automated PII detection and classification across Bronze, Silver and Gold.
  • Controls access: role-based grants plus attribute-based policies, row filters and column masks (null, redact, partial, hash, year only).
  • Governs AI: Voir agents are scoped principals; masks, row filters and PII guardrails apply to every tool call and answer.
  • Proves it: lineage from source to metric to decision, and an audit log of who or what accessed which data.
  • Stays local: Atlas runs inside your deployment, so policies, lineage and logs never leave your jurisdiction.

These controls support GDPR, PIPEDA, Quebec Law 25, CCPA, HIPAA and sector rules such as OSFI E-23. Read DataNature Atlas in depth or see security and compliance.

See it on your own data.

Book a demo

Demos and quotes are handled by Dlytica, the company behind DataNature.