Problem
Account takeover (ATO) is a fraud risk in online payments: a fraudster gets into a real customer's account. At PayPal one pattern that frequently appears in taken-over accounts, repeated transfers between two accounts, causes millions in fraud losses every year. Individual events look ordinary, so the signal sits in the links between accounts: shared devices and addresses, repeated transfers, tightly connected groups. In a relational store these links need several self-joins of the transactions table and heavy compute.
Example from the article: two accounts, A and B, send money back and forth at least four times in a period (the "ABABA" pattern), which appears often in taken-over accounts.
Solution in brief
Store accounts, assets (IP, address, VID, device ID) and transactions in a real-time graph database. Three signals then become simple graph queries: how many accounts share an asset, whether a known transaction pattern such as ABABA occurs, and which connected communities look unlike the rest of the graph.
Questions this case answers
- How to detect coordinated abuse by accounts when single-entity signals look normal and the evidence is in shared attributes and repeated interactions between entities
- How to query multi-hop interaction patterns across a large event log when repeated self-joins of a relational table are too expensive
- How to detect account takeover (ATO) when each signal looks normal on its own and only the links between accounts give it away?
- How to find accounts that share an unusual number of addresses, devices or IPs?
Tools and methods
- Asset-account graph — accounts linked to assets they use; finds accounts with an abnormal number of shared assets
- Transaction pattern matching — extracts repeated shapes such as ABABA from the account-transaction graph
- Graph features (connected community) — groups of linked accounts for clustering and anomaly detection
- Real-time graph database — stores and queries the graph
Google or a one-time email code. No password.