A glimpse into the cutting edge of anti-fraud research

Jia Chen, Head of Data Science, Integrity & Core Experience; Wei Yang Wang, Data Science Manager

. October 24, 2024 . Regional , Singapore

Digital fraud is a significant risk for Grab. An ecosystem like ours—with millions of daily transactions between merchant-, driver- and delivery partners, and consumers—is a prime target for bad actors.

Common fraud attempts we encounter include account takeovers, where fraudsters seek access to legitimate Grab user accounts through phishing or social engineering. Another example is credit card fraud, where criminals use stolen credit card information to transact on the platform.

Grab’s Integrity team has been developing its own anti-fraud technology to combat fraud. Our enterprise service, GrabDefence, not only protects Grab’s platform, but is also used by other fintechs in the region, such as Vietnamese mobile wallet Momo and Indonesian P2P lender Julo.

(Also read: GrabDefence: Building anti-fraud technology for Southeast Asia’s digital age)

However, digital fraud techniques keep evolving. While tools to detect, prevent, and manage fraud are getting more sophisticated, this can’t be done by one industry player alone. It requires long-term efforts involving multiple stakeholders, including practitioners like Grab, and academics.

Pushing boundaries is a team effort

That’s why a couple of us at Grab’s Data Science team are working closely with researchers at the National University of Singapore (NUS) to make breakthroughs in digital fraud research.

This effort is funded through a multi-year government-backed research programme called AI Singapore (AI SG). Our main research focus under AI SG is to harness the potential of graph-based models and deep learning.

Because digital fraud affects all of us, we want to share a little on the work we do with AI SG to counter it.

Graph-based fraud detection

Let’s start with a key concept: What are graphs, in the fraud research context?

You can imagine a graph as a data structure where we connect everything—all the individual data points of users, merchants, partners, and transactions. In contrast, traditional data structures such as tables put all these separate entities into different rows of data, thereby missing out on important connectivity information.

They are a helpful tool in fraud detection, as they can help us visualise and analyse complex relationships between data points. We deploy deep learning, an artificial intelligence methodology which is particularly good at making sense of large amounts of unstructured data, such as those represented in these graphs.

With deep learning models such as graph neural networks (GNNs), we can capture more fraudsters who would have otherwise slipped under the radar.

Take credit card fraud as an example. Graphs can help identify potential fraudsters who repeatedly misuse stolen credit card information across numerous fake accounts. Similarly, in the case of money laundering, graphs can reveal suspicious connections where money is funnelled through multiple financial entities repeatedly in order to hide the source of funds.

With GNNs, instead of examining each connection individually, we can simultaneously consider how connections are related to each other and how they interact. This helps to identify fraud rings more effectively.

(Also read: Graph Networks – Striking fraud syndicates in the dark)

Taking graph analysis further

The next stage is where our work with NUS and AI SG comes in—we share the same ambition to improve and innovate the way we work with graphs and GNNs.

As an industry player, we emphasise time-to-market and real-world impact, while in our research work with universities we can strive for technical excellence on longer time frames and with more exploratory approaches.

One target of our long-term research is improving the robustness, efficiency, and scalability of existing models.

For instance, we developed a technique called SPADE+ which can better handle evolving graphs than previous models. Transaction data on digital platforms like Grab isn’t static; it keeps changing in real-time.

To make sense of it effectively, SPADE+ offers ways of not just adding new data, but deleting old data in a smart way. Smart, in this context, means adding and subtracting data adaptively in batches only when potentially fraudulent transactions are detected. This helps to improve overall latency.

In a reality where successful fraud attempts can cause substantial financial and reputational damage, the industry requires fraud detection systems capable of acting within hundreds of milliseconds on million-scale graphs.

Since the AI SG funding was granted in 2022, we’ve co-authored multiple academic papers, including SPADE+, and presented at prestigious conferences in the machine learning community such as the International Conference on Learning Representations (ICLR) and the International Conference on Very Large Data Bases (VLDB), benefiting the larger scientific community.

We have regular check-ins, provide ideas and input to the research team, and serve as a test bed for them to validate their hypotheses.

Academic research takes time, which is precisely why we need it. We intend to incorporate some of the models for use in Grab once the research matures.

Taking a collaborative approach helps achieve impact in a deep and meaningful way, on things that truly matter to the partners we serve.

Industry trends

A glimpse into the cutting edge of anti-fraud research

Pushing boundaries is a team effort

Graph-based fraud detection

Taking graph analysis further

Komsan Chiyadis

Komsan Chiyadis