Skip to Main Content Skip to Footer Content
Close Search

Reading The Signals Right: Why Network Intelligence Is The Future Of Machine Learning For Payments Fraud Prevention

Why Network Intelligence Is The Future Of Machine Learning

Today’s payments ecosystem is more open and complex than ever. New payment methods and financial products are quick to market, digital personas have become the new consumers and the speed of payments has increased significantly with accelerated digitalization. This means more opportunities for fraudsters. Financial institutions (FIs) must therefore assess risk in real time – and more efficiently than ever – but their heavily invested machine learning (ML) solutions may be challenged with a lack of history from which to learn.

ACI experts Cleber Martins, head of payments fraud solutions for banks and intermediaries, and Damon Madden, principal fraud consultant, explain what this landscape means for the future of machine learning technology and why adopting a network intelligence approach can help FIs stay ahead in the fight against fraud.

Chris Taine: Can you describe today’s payments ecosystem and some of the challenges it presents for fraud detection?

Cleber Martins: The payment ecosystem is multi-player, fast, interoperable and connected. With increased granularity, payment players in their siloes cannot see the overall picture as they react to fraud attempts. Cybercriminals, who are increasingly connected and organized, can impersonate consumers’ digital personas from anywhere. They can perpetrate fraud throughout this ecosystem, while single players are limited by what they know or have time to build. This represents rich pickings for fraudsters.

CT: And why is this a challenge for machine learning solutions?

CM: Machine learning (ML) is still the right solution, but it is only as good as the signals from which it learns. A predictive machine learning solution will learn from historical data, and the strongest signals for predicting fraud come from understanding each persona. But, if there is no profile for a new digital persona, there is neither time nor opportunity to build it before the fraud takes place. Custom approaches for ML have led to recent investments exactly because of the stronger signals that an FI could get from its own profiles, products, offers and segments.

The traditional consortium-type ML approach to collaboration comes with too many trade-offs. In addition to privacy and regulatory constraints, it can miss the value of the FI’s individual signals, which is important given each FI should maintain its individuality.

CT: What are these trade-offs?

Damon Madden: For one thing, the “one-size-fits-all” view of the consortium means the bigger members disproportionately influence the group’s perspective — despite the smaller players usually facing fraud first. But what really works against the consortium is the slow turnaround — collecting data, building it into a usable model and distributing it to the rest of the group. Relative to the lightning-fast requirements of the digital age, this approach is no longer fit-for-purpose, not to mention that privacy constraints limit the approach significantly. The new financial ecosystem Cleber described needs a new fraud protection ecosystem. Network intelligence provides that.

CT: So how is the network intelligence approach different and better?

CM: Nobody knows as much as everybody. Instead of being in a bubble, an FI becomes part of an intelligence community that facilitates consuming, exchanging and distributing risk signals. It really enables a hybrid-type ML approach that leverages the best of one another’s approaches. FIs can maintain the strength of their custom-type signals, complemented by signals exchanged with their community as well as signals from third-party fraud intelligence sources. Network intelligence connects members in the middle of their ML process, so no raw data sharing is needed, and signals are anonymized to preserve FIs and their customers. The actual ML outcome is tuned for each individual FI. Making precise risk assessments in real time about a digital persona without knowing its reputation across the ecosystem is otherwise not possible.

CT: So how would a typical transaction look under this approach?

DM: Signals are ML-specific language, built by mathematical algorithms that express risk relevance and reliability in metadata form that ML can understand and correlate with other signals. The result is more precise risk assessment. The signals need to be available for an ML solution to use: They can be pre-calculated and made available via metadata packages, stored in ML signals libraries ahead of time; provided via an API call in real time, or they can be calculated in-flight. The most relevant scenario is for those risk signals to accompany the transaction itself. In a scenario where a transaction passes through various participants within the payments ecosystem, each member can add signals to the transaction before passing to the next member. At the end of the cycle, the process will form a true DNA of risk, which uniquely expresses the many different perspectives of risk as perceived by the community.

The power comes from the signals that give each FI knowledge that they would not have on their own, ultimately protecting the entire ecosystem. Signals do not need to identify the source or even what would make it meaningful for a human; they empower the ML solution to find the most precise correlation to stop even the most complex fraud attempts, anonymously. That’s the power of ML, so network intelligence makes ML stronger than ever before. A member of the network intelligence community is connected to one or multiple hosts, so smaller communities can operate totally independently, larger communities can be formed, and there is no real limitation to the use cases that can leverage this technology and methodology to become stronger. Once in the network, members have immediate value from the next piece of risk intelligence available.

CT: What is a “network intelligence host” and what role do they play in this approach?

DM: A network intelligence host is a really novel aspect of this approach and crucial to the collaboration process. They are network members —a processor, a telco, a data bureau, some form of central infrastructure or anyone that has intelligence to enrich fraud prevention — that bring in-depth knowledge of a particular discipline to the table. That knowledge is converted into signals that provide our machine-learning models with better insights, enabling more precise decision making.

Take ACI as an example. We’re one of the largest eCommerce fraud prevention businesses, so as a network intelligence host, we enable everyone on the network to leverage the signals we build from our database and the work we do with merchants. We also help the merchants in return, by stopping fraudulent transactions before reaching the FIs, not only reducing losses, but also operational costs across the ecosystem.

CT: But how easy is it for members to integrate this approach with what they already have in place?

CM: The beauty is that the approach complements existing fraud solutions. Members of the network have a choice in what signals to adopt, how to apply them, and are informed on the relevance and reliability those signals are expected to bring to their portfolios. FIs can develop their own signals, embrace new ones in their current strategies, or build new ML strategies in mere hours (if not minutes). It’s important to note that the entire solution was developed from a business user perspective. Network intelligence and the entire methodology we’ve developed gives business users full control of their business results. This is the real democratization of machine learning, where we get the technology of artificial intelligence to work for the business, and not vice-versa. 

For more on ACI’s network intelligence concept, listen again to Cleber and Damon’s deep dive in our recent webinar, Leveling Up Your Fraud Detection with Network Intelligence. You can also learn more by reading the accompanying eBook, Expanding the Horizons of Fraud Detection.