Hungry AI: How Machine Learning Technology Is Transforming Payments
2018 was a busy year for the payments providers.
For example, last year PayPal — one of the premier global payment processors — went on what could be considered a shopping spree. In Q2 the payments giant made four major acquisitions totaling roughly $2.7 billion dollars. They purchased iZettle to expand their small business footprint in Europe, Jetlore to deepen their personalization capabilities, Hyperwallet to expand their marketplace payments strategy, and Simility to enhance their anti-fraud suite.
Elsewhere, Stripe deepened their value proposition, Ayden went public, Visa and Mastercard got behind Secure Remote Commerce (SRC), Vantiv acquired Worldpay, and Zelle expanded in its first year live.
This explosion of activity in the payments market leads to lower costs, easier options for consumers, better mobile and wireless options, and vastly more transactions. And wherever you have transactions, you have payment fraud — in fact, wherever you have money, you have fraud. Industry rule of thumb for card-not-present payment transactions has been 4% of the dollar value is fraudulent — a number which could swing up or down depending on whether countermeasures or thieves have the upper hand.
With fraud so widespread across the payments sector, it’s no wonder companies are scrambling for tools to fight it — and maintain a good user experience, not questioning and requiring verification for every valid transaction.
But fraud for payments is nothing new.
Past systems used “decision engines” where deterministic algorithms were generated by back-end data scientists, and often refreshed on a month-by-month basis. A fun book from this period is Kevin Poulsen’s book Kingpin, describing forging credit cards and exploiting them. Well-known triggers would be someone buying four identical high priced watches, or simultaneous transactions at great geographical distances.
The current battle is fought with bringing more data to bear, but also bringing online new filters to counter new threats at a moments notice — far beyond simply executing static rules. However, most existing data infrastructures aren’t built to handle real-time data analysis at the global scale needed to mitigate the risk of fraud.
It is unlikely your data scientists will be watching at 3 am to counter a new attack, and unlikely they can finish an analysis during the peak of a threat. Like a Hollywood heist movie, the police show up to find an empty vault and the money long gone.
AI, however, can provide a missing link of real-time flexibility, and the ability to respond in moments, if executed correctly, giving time for the quantitative teams to spring into action. “Hungry AI,” in particular, with access to broad swaths of data and using automatic feature selection mechanisms, can provide the effective short term response, and also help your analysts detect subtle patterns.
Here’s what it is and how businesses can use it to stop bad actors:
What is “Hungry AI”?
The team at Aerospike coined the phrase “Hungry AI” to popularize the idea of large-scale AI, with literally hundreds of billions of data points, where the models have a broad range of features available simultaneously.
Today, many companies use predictive analytics to enhance productivity in areas like marketing campaign optimization, risk assessment, market analysis, and fraud detection. But with traditional techniques, it’s often hard to predict how much data a particular model will effectively need. That’s where your data scientists attempt to select the correct AI models, and to protect against overfit and similar problems.
However, due to front-edge database realities, they are often forced to compromise.
Hungry AI enables you to engage in meta-optimization to discover the best possible set of parameter values for a given machine algorithm — and to allow your higher level AI feature selection to respond in real-time, changing those parameters, perhaps bringing into play a dataset that was not previously effective.
Having those extra features online requires hundreds of terabytes and potentially ingesting extra hundreds of billions of events per day. Few people believe databases exist which have the efficiency to store and retrieve that quantity of data and manage it to enterprise-grade standards, with millisecond-level response times.
The reality is Aerospike is being used — live — in multiple use cases with major payment processors today. In exactly this way, providing a reasonable cost solution with far higher levels of performance and reliability than systems like Cassandra and Hbase.
If you aren’t using this kind of system, you might be the kind of “soft target” an attacker will focus on, to the peril of your business.
An example — ThreatMetrix
ThreatMetrix is an identity management company. They ingest massive internet data sources, including clickstreams and advertising based flows, to validate whether someone is who they say they are. They do this as a service, bringing together multiple real-time data sources in their data centers.
A large payment processor or credit card company will ask ThreatMetrix to verify a buyer’s credentials, having provided recent behavior. At this point, a combination of algorithms, both AI and deterministic, spring into play. Within milliseconds thousands of datapoints are considered, calculations are done, and the buyer is either validated, or the payment processor is told to apply extra scrutiny to the transaction.
When there’s a new threat, they’re able to respond to an attack very quickly.
Hungry AI quickly allows them to keep more data online, to ingest data with sub-second delay, and thus to catch the bad actors while at the same time avoid denying payments that should go through.
The more data you have, the better Hungry AI works.
Today, large payments and credit card processors have a wealth of data which they could bring to bear. Gigabyte datasets can be collected by smaller players, but the larger companies can apply their larger market footprint to great positive effect.
With AI, the greater amount of data can be used more effectively. Models can retrain on a moment-by-moment basis. As models are exposed to new data, they are able to independently adapt — they learn from previous computations to produce reliable, repeatable decisions and results.
Don’t be the soft target.
