Curve has adopted Google Cloud's BigQuery Graph to identify linked fraud networks across its payments platform.
The move addresses a recurring challenge in fraud detection: fraudsters frequently share attributes across multiple accounts, including devices, funding cards and contact details, creating connections that are difficult to trace using conventional relational databases.
In a standard relational database, uncovering these hidden links requires analysts to run repeated self-joins to follow chains of connections between accounts. According to Curve, this process became increasingly expensive and difficult to maintain at the scale of millions of users and tens of millions of connections. Some of the company's more detailed fraud signals involve billions of possible links, which created further performance constraints within standard relational systems.
To resolve this, Curve modelled its payments environment as a graph, representing users as nodes and shared identifiers as edges. Analysts can now search for suspicious patterns across the dataset using graph query language, while keeping data within its existing BigQuery warehouse rather than migrating to a separate graph database. In addition, Curve said this approach reduced both the time and cost typically associated with such a migration and allowed teams to combine graph traversals with SQL analysis and machine learning workflows in the same environment. Existing SQL pipelines continue to build the underlying nodes and edges tables, with graph queries used to traverse relationships and SQL applied again for aggregation.
Financial and operational impact
Curve said automated blocks triggered by graph-based analysis prevented approximately USD 12 million in transaction losses during 2025, with graph-based queries reaching around 72% accuracy in identifying fraudulent users. That level of precision has allowed fraud mitigation staff to focus manual reviews on higher-risk cases, while graph query language has made it easier to update fraud rules more frequently. Previously, hourly rules were limited to one-hop queries, as more complex scripts were too slow to run efficiently. The newer approach has enabled Curve to expand its search for broader account and device networks linked to organised fraud activity.
Next steps: real-time detection and visualisation
According to the announcement, the company said faster graph traversal also has implications for the machine learning models used in its fraud monitoring. While daily graph rebuilding and traversal are adequate for training models, they are too slow for live transaction decisions, which may need to be made in under a second. As a result, the company is moving towards micro-batch or streaming graph traversals to feed fresher network data into fraud models during live monitoring, and is working to incorporate higher-volume signals, including billions of IP address connections, into real-time detection processes.
Curve is also reviewing graph visualisation tools that would allow security and data science teams to inspect emerging fraud networks visually. The company said the broader shift reflects a move towards treating fraud detection as network analysis rather than isolated transaction review, while reducing the operational burden associated with complex relational queries.
