Coinbase upgrades anti-fraud system: integrating machine learning and rule engines to reduce response time to several hours

Odaily Planet Daily reports that Coinbase says it is optimizing the rule creation process in anti-fraud systems by integrating machine learning models with rule engines, achieving more efficient risk management. It also proposes a dual-track strategy of “models responsible for long-term defense, rules responsible for rapid response,” and builds a unified framework so that the two form a feedback closed loop: rules are used to catch new types of fraud and, in turn, train the models inversely, thereby continuously improving overall defense capabilities.

For specific optimizations, Coinbase is turning the previously manual rule creation process into data-driven workflows with automatic recommendations by restructuring data structures, automating Schema evolution, and introducing notebook-based analysis tools. This significantly improves efficiency. Among them, rule backtesting performance has increased by more than 10 times, and overall response time has been reduced from several days to a few hours. In addition, the new system uses machine learning to recommend parameters, which helps reduce false positives and, while combating fraud, also reduces the impact on normal users.

Coinbase says that the next step will be to advance event-driven automatic rule generation, and to explore converting high-efficiency rules—one click at a time—into model features, further moving toward an automated risk management system.