Real-time deep learning fraud detection networks for e-commerce platforms

Grab's ecosystem processes millions of daily transactions among merchants, drivers, delivery partners and consumers, drawing constant fraud pressure. Common attacks include account takeovers via phishing or social engineering and credit card fraud. Grab already used graph models to surface suspicious linkages, then deepened the work through AISG and NUS, focusing on graph models and deep learning.

Graphs represent entities as nodes and connections as edges to expose hidden patterns. Graph neural networks improve detection by learning from both node properties and connection structure. Spade detects fraud in evolving graphs; Spade Plus adds user-defined functions and stronger streaming throughput. Rush catches sudden fraud spikes and PMP handles diverse connection types. To process the largest networks, the team built dynamic random walk on FPGA hardware — a high-speed scanner that catches and removes fake users in real time. Since 2022 the collaboration has produced multiple academic papers; mature models will be incorporated into Grab production.

Caption language: en · Fetched: 2026-05-02

In today's fast-paced digital world, e-arketplaces are growing at an unprecedented rate. Millions of transactions occur every second. But hidden among them are patterns that suggest fraud. Fraud on digital platform leads to financial losses, damaged reputations and also higher security cost for industry and everyone. With millions of transactions within Grabs ecosystem between our merchant driver and delivery partners and consumers, we are a prime target for bad actors. Common fraud attempts include account takeovers through fishing or social engineering and credit card fraud. At N US, we've developed AI powered solutions to tackle fraud in dynamic, large scale, and fastmoving environments. Grab has been leveraging graph models to detect fraud.

Graph models enable us to catch fraud more effectively by leveraging linkages through suspicious connections. The collaboration with N US through AI Singapore focusing on graph models and deep learning help us to deepen our technology in fraud detections. Gras is very of powerful data structure representing entities as nodes and their connection as edge helping to uncover hidden patterns in the data. So graph neuronet network improve the detection by learning from both property in each entity as well as how they are connected. Spay detects the flot in unvolving graphs while Space Plus enhance it even further with userdefined functions and better streaming efficiency. Rush catches sudden fraud spikes fast while Spade Plus tracks evolving patterns together forming a powerful responsive defense system.

Graph-based fraud detections face challenge like label imbalance, changing network and increasing powerful fraud sticks. To solve this, we propose PMP to handling diverse connection spade for realtime updates and also powerful G techniques to detect hidden frauds in the data. Like a student practicing varied problems, AI learns better with diverse data, helping it catch evolving fraud even as behaviors change. We've developed techniques to better handle evolving graphs and adapt to real-time data changes. We've also explored anomaly detection strategies that help enhance performance in identifying unusual patterns across various graphs. Amidst the large pool of data, we have found a way to understand big networks using special hardware, FPGA.

Dynamic random walk on FPGA is a super fast detective that scans social networks to catch and remove fake users in real time. Since 2022, we've co-authored multiple academic papers and pre presented at prestigious conferences. We intend to incorporate some of the models for use in GRAP once the research mergers. In the future, AI will not only react but also predict and prevent fraud before it happens. Fraud is constantly evolving, but so are we. With AIdriven fraud detection, we're securing the future of earketplaces, one transaction at a time. [Music]