AusPayNet’s most recent fraud statistics saw a 15.4 per cent decline in payment card transaction fraud for the year to June 2020, continuing a downward trend that started the year prior.
It’s the result of a coordinated industry effort in Australia to combat fraud associated with online transactions. Until that effort, online fraud had grown roughly in line with the rise of e-commerce.
But there won’t be celebrations just yet.
For starters, online shopping burgeoned last year. Australia Post, which sees the logistics and last-mile delivery, reports that e-commerce grew 57 per cent year-on-year in 2020. As of May this year, online shopping was still 37 per cent higher than a year ago.
We don’t yet know how this has impacted fraud levels. As the Australian government recently noted, “industry initiatives have created a general decline in card fraud, making the actual effect of the pandemic difficult to quantify in the short-term.” Card fraud, while being reduced, is still a $447 million problem. Fraudsters are “relentless” and won’t disappear anytime soon.
In short, there’s still room for concerted action on fraud.
Australia has already tasted success through targeted action. Throwing machine intelligence and more powerful compute capacity at the problem may allow Australia to consolidate and build on its gains, flushing out additional, otherwise hard-to-detect, fraud instances.
The real-time march
In finance, there’s a general emphasis on moving to real-time or near real-time processing across a range of functions, both customer facing and behind the scenes.
Australia’s financial sector is undergoing a step change away from batch processing of transactions to more straight-through and real-time processing. These newer payment architectures are digital from end-to-end, and offer banking customers much faster ways to pay. There’s no doubt this represents the future of payments.
Similar trends are occurring in the way transactions are screened for fraud. Historically, fraud detection was rule based – for example, if a single credit card is used on three websites in under 10 seconds, then lock the account and notify the owner. As fraudsters learned to game these systems, companies turned to computers – and specifically machine learning models – to help.
Machine learning is now a vital tool in fraud detection. Powerful algorithms can quickly process millions of data points and make connections between unrelated data sets in order to detect suspicious patterns – at a much greater speed, scale, and efficiency than rule-based systems.
Companies use machine learning models to ensure that their customers are who they say they are and that neither the customer or the business is being defrauded. They also run these checks in as close to real-time as possible. For a lot of the card processing companies, their window for authorisation of a transaction is 40 to 50 milliseconds. As EY recently noted, “most banks [in Oceania] are already performing near real-time fraud checks on payments as standard".
The need for vigilance
But near real-time is not the same as real-time. As more transactions occur in real-time, fraud detection must follow suit. That will require a review of the infrastructure used to underpin that detection capability.
Machine learning for fraud detection requires stream processing in order to make calculations in real-time and cutting-edge in-memory data stores to provide accelerated model development and performance.
High volumes of incoming transaction data need to be scored against the models. The largest retailers may process tens of thousands of transactions a minute, and every one of those needs to be checked for potential fraud.
This data needs to be scored in real-time. In addition, fraud detection systems need to be able to talk to other systems, such as a payment processing system, preferably in real-time.
At the volume and speed that is necessary, only in-memory technologies will suffice. In-memory technologies are designed to deliver the necessary performance to make this possible.
It isn’t just the growth of e-commerce volumes and real-time payments that will drive the need to improve fraud detection systems.
As fraudsters evolve and fine-tune their methods, companies need to be ready.
But more than that, as billions more are spent by Australians shopping online, merchants want to provide the most frictionless e-commerce experience. Research shows at least 50 possible points of friction already in the end-to-end e-commerce experience.
Fraud detection systems need to be fast and unobtrusive. They need to work quickly to verify a customer at sign-up, and distinguish fraud from irregular but legitimate activity. They shouldn’t add to e-tailers problems.
For example, by recognising patterns in customer behaviour through machine learning, a system may identify small changes that are uncharacteristic of a customer, use those as flags for potential identity theft, and send a verification request to the customer. Consumers typically appreciate this kind of unobtrusive fraud protection, and it can increase trust with the merchant.
John DesJardins, chief technology officer, Hazelcast