Production Architecture: Fast Path vs Slow Path for Adversarial Defense

Two-Path Architecture

Production fraud systems use parallel detection paths. The fast path runs lightweight rules and models under strict latency constraints (10-50ms)—it must make a decision for every transaction in real-time. The slow path runs more sophisticated adversarial-robust models asynchronously, analyzing transactions after the fact and feeding signals back into the fast path.

Architecture Pattern: Fast path: rules + lightweight ML (immediate decision). Slow path: adversarial-robust models, graph analysis, behavioral clustering (delayed enrichment). Slow path findings update fast path thresholds and blocklists.

Why Two Paths

Adversarial-robust models are computationally expensive—ensemble methods, multiple feature representations, deeper networks. Running these on every transaction in real-time is cost-prohibitive. The slow path processes transactions in batch, using expensive analysis only where the fast path flagged uncertainty or after suspicious patterns emerge.

Fast Path Design

Optimize for latency and false negative rate. Use velocity checks, blocklists, and lightweight models. Accept some false positives—the slow path can correct them. Design fallback behavior: if fast path times out, default to allow (with slow path review) or block (conservative) depending on risk tolerance.

Feedback Loop: Slow path identifies new attack patterns. These become new fast path rules or model features. Continuous deployment updates fast path defenses within hours of attack detection.

Slow Path Capabilities

Graph analysis across transactions (detecting coordinated attacks), behavioral anomaly detection over longer time windows, ensemble voting across multiple model architectures, human review queue for edge cases. Latency budget: minutes to hours.