Zürcher Kantonalbank Deploys AI-Powered Fraud Detection

Title

Zürcher Kantonalbank Deploys AI-Powered Fraud Detection

Content

Zürcher Kantonalbank (ZKB), Switzerland's largest cantonal bank and a systemically important institution, has deployed an AI-powered fraud detection platform using deep learning and real-time behavioural analytics across digital banking, payments, and securities trading — building on a long-running partnership with Swiss fraud-prevention specialist NetGuardians that went live in 2021 and has since been tuned to address deepfake identity fraud, social engineering attacks, and automated account takeover schemes facing Swiss banks. The platform was built because traditional rule-based systems are increasingly insufficient against adversaries who adapt to circumvent static rules. ZKB processes millions of transactions daily across retail banking, corporate banking, payment services, and securities trading — both a significant fraud exposure and an extensive data foundation for training models. The platform builds dynamic behavioural profiles from each transaction and customer interaction, enabling detection of novel fraud types that rule-based systems would miss and significantly reducing false-positive rates that have historically created friction in legitimate customer transactions. The architecture comprises deep learning neural networks trained on historical transaction data and enriched with external fraud intelligence feeds and synthetic fraud scenarios generated through adversarial modelling. Models operate at multiple temporal scales, analysing individual transactions while tracking customer behaviour over hours, days, and weeks to identify trend changes that may indicate account compromise or gradual fraud. Graph neural networks analyse network relationships between accounts, counterparties, and transaction flows to identify money laundering networks, fraud rings, and account takeover cascades. Behavioural biometric analysis creates dynamic profiles of each customer's interaction patterns — typing, mouse movement, navigation, device fingerprints, session timing — generating risk scores that can trigger additional authentication, transaction holds, or alert escalation. The system processes transactions within milliseconds using a tiered decision framework: lightweight models screen all transactions (approximately 95% pass through without delay), the remaining tier applies deep learning, graph analysis, and behavioural assessments, and high-risk transactions escalate for manual review. The platform integrates with the bank's anti-money-laundering framework and extends AI-driven surveillance into securities trading to detect spoofing, layering, and wash-trading patterns. Future priorities include deeper integration between fraud detection and AML systems, expansion to additional transaction types, and continued adaptation to emerging threats in the Swiss and European financial crime landscape.