Coinbase upgrades its anti-fraud system, integrating machine learning with a rules engine, reducing response time to a few hours

Coinbase stated that it is optimizing the rule creation process in its anti-fraud system by integrating machine learning models with a rules engine, achieving more efficient risk management. It also proposed a dual-track strategy of "models responsible for long-term defense, rules responsible for rapid response," and built a unified framework to create a feedback loop between the two: rules are used to capture new types of fraud and train the model in reverse, thereby continuously enhancing overall defense capabilities.

In terms of specific optimizations, Coinbase has transformed the previously manual rule creation process into a data-driven and automated recommendation system by restructuring data, automating schema evolution, and introducing notebook-based analytical tools, significantly improving efficiency. Among these improvements, the performance of rule backtesting has increased by more than 10 times, and the overall response time has been reduced from several days to a few hours. Additionally, the new system uses machine learning to recommend parameters, helping to reduce false positive rates while combating fraud and minimizing the impact on normal users. Coinbase indicated that the next step will be to advance event-driven automatic rule generation and explore the "one-click conversion" of efficient rules into model features, further moving towards an automated risk management system.