Plaid introduced a transformer-based Fraud Foundation Model for Plaid Protect, pretrained on behavioral and transactional sequence data from the Plaid Network.
Architecture: transformer attention mechanism weighing events across sequences; pretrained broadly then adapted to fraud tasks rather than rebuilt per use case.
Training: hundreds of millions of proprietary data points, fine-tuned on customer-specific fraud labels. Feeds into existing Trust Index scoring framework.
Internal tests claim up to 40% improvement in fraud detection performance over previous Trust Index models. All metrics are vendor-reported.