Fraud detection balances security and friction:
Rule-based checks:
- Velocity limits (too many transactions)
- Geographic anomalies (purchase from new country)
- Amount thresholds
- Known bad actors (IP, device, card)
ML models:
- Train on historical fraud labels
- Features: transaction patterns, device fingerprint, behavioral signals
- Real-time scoring on each transaction
Response actions:
- Allow: Low risk
- Challenge: DS, SMS verification
- Block: High risk, decline transaction
Trade-off: Too strict blocks legitimate users. Too loose loses money. Optimize for total revenue.