Learn→Fraud Detection & Anomaly Detection→Supervised Anomaly Detection (Imbalanced Classification)→6 of 6

Fraud Detection & Anomaly Detection • Supervised Anomaly Detection (Imbalanced Classification)Hard⏱️ ~3 min

Failure Modes: Concept Drift, Adversarial Attacks, and Cold Start

Concept Drift
Fraudsters adapt. When models learn to catch a pattern, fraudsters change tactics. Last month's attack vector becomes ineffective; new vectors emerge. A model trained on historical data degrades as the fraud landscape evolves.
Concept drift in fraud differs from natural drift. It is adversarial: fraudsters actively probe model boundaries and exploit weaknesses. Feature importance shifts as fraudsters discover which signals get them blocked and which do not. A feature that was highly predictive becomes useless once fraudsters learn to avoid it.
Adversarial Attacks
Fraudsters conduct probe attacks: small test transactions to map model behavior. They vary one feature at a time to identify thresholds. "How much can I spend before getting blocked? Which device fingerprints pass? What velocity triggers review?"
Defense requires obfuscation and non-determinism. Add random delays before blocking high-risk transactions so timing does not reveal threshold hits. Use randomized thresholds within a range. Limit the information leaked through block messages.
Cold Start
New accounts, new merchants, new devices have no history. The model relies on aggregate features (velocity, spend patterns) that do not exist yet. First-party fraud often uses newly created accounts specifically to avoid historical signals.
⚠️ Cold Start Mitigations: Use graph features connecting new entities to known ones (shared device with flagged account). Apply stricter thresholds for new entities. Require additional verification (SMS, email) for first transactions.
Monitoring for Failure
Track feature importance over time. If a top feature drops in importance rapidly, fraudsters may have adapted. Monitor false negative rate by segment: a sudden spike in chargebacks from a specific merchant category or device type indicates a new attack pattern the model misses.

💡 Key Takeaways

✓Concept drift is adversarial: fraudsters actively probe and adapt to exploit model weaknesses

✓Probe attacks map model behavior by varying one feature at a time to identify thresholds

✓Defense: add random delays, randomized thresholds, limit information in block messages

✓Cold start: new entities have no history; use graph features linking to known entities

✓Monitor feature importance drops and chargeback spikes by segment to detect new attack patterns

📌 Interview Tips

1Explain adversarial drift: fraudsters discover which signals get blocked and avoid them

2Describe probe attacks: test transactions vary one feature to map thresholds

3For cold start, use graph features: does new account share device with flagged account?

← Back to Supervised Anomaly Detection (Imbalanced Classification) Overview