Mimicking fraudsters

Credit card fraud losses in the UK fell for the first time in nearly a decade last year, by more than 5% to £402.4m, according to research by the Association of Payment Clearing Services (Apacs).

The fall has put a spotlight on the increasing use of neural networks that have the ability to detect fraudulent behaviour by analysing transactions and alerting staff to suspicious activity. As commercial applications of research into artificial intelligence, these systems give the impression of mimicking human abilities for recognising unusual activity.

Karina Purang, a financial analyst at Datamonitor in London, says the use of neural networks is growing: "These systems are very important to banks trying to reduce fraud, and are becoming standard across the card industry to detect unusual spending patterns." She says Barclays reported that after installing Fair Isaac's Falcon Fraud Manager system in 1997, fraud was reduced by 30% by 2003. The bank attributed this mainly to the new system.

Neural network software systems are integrated with a bank's card management and authorisation systems. The software consists of a range of programs that recognise patterns in cardholder behaviour - the fruits of a long history of development work into using computers for pattern recognition. The system monitors cardholder behaviour, looking for unusual transaction volumes, amounts and locations. It also looks at the types of merchants being used, and even the day of the week, for things that don't fit a pattern based on an individual's credit card history.

The system scores each transaction, giving higher scores for those it suspects could be fraudulent. Banks either monitor the results in real time or analyse periodic reports.

Banks buy or lease the software from a few specialist firms that operate globally, giving them the ability to compare data about fraud patterns in different countries.

"It's a huge challenge because you are trying to detect the fraud that occurs in around one in 50,000 transactions," says Sushimito Ghosh, executive vice president of Retail Decisions, a provider of neural network systems. "The other problem is that fraudsters are constantly modifying their behaviour as soon as one kind of activity is detected."

Martin Hewgill, a senior business consultant at ACI Worldwide, says that typically, a neural network needs six months of recent credit card transaction data to allow it to "learn" patterns of fraud. "From this, it is able to build profiles and detect clumps of activity that show a probability of being fraudulent."

But he also admits that fraudsters modify their behaviour as detection gets better. "When they find out a bank has such a system, they target other banks or use other techniques. That means that over an 18-month period, the system's value declines as it becomes less able to detect new types of fraud."

Experts also say there is a debate about the best way to use the scores that systems deliver. "Do you just target high scores or take action on lower scores in an effort to detect more fraud? We favour a mixed approach," says Ghosh.

In reality, it is the bank's fraud team that has to make the decisions. The team uses the alert data from the system to apply the bank's rules and policies on how to react to a suspected fraud. If the concern is great enough, someone usually calls the cardholder to verify the transaction. The next level of escalation is to stop the transaction.

The ability to reduce the fraud level is always a trade-off between the cost of installing and running the system and the amount of fraud taking place. Cheaper neural network systems make it economical to reduce the level of fraud.

Today's neural network developers are competing to deliver systems that are faster, learn more quickly, and are easier to use, because they provide better links to the fraud management systems that banks use to evaluate alerts.

Nick Sandall, head of retail banking at Deloitte, says that banks also use other technologies. "The artificial intelligence community is constantly bringing us new solutions. These include hybrid systems that build up and analyse patterns of transactions. But the big problem with credit cards is you have milliseconds to decide if it's a fraudulent transaction, otherwise it is analysis after the event."

AI systems will be expected to look more closely at online and telephone transactions, following the introduction of Chip and Pin.

Chip and Pin - which has been very successful in France - is now being rolled out in the UK, with pin number entry pads gradually replacing signature verification in retail outlets. However, banking experts predict that, as a result, fraudsters will increasingly target other areas, such as online payments.

Further into the future, neural networks could be replaced or complemented by systems that mimic biological patterns. New research into computational immunology is looking at fraud as if it were an attack on a healthy system.

But experts agree there is always a trade-off between the cost of implementation and the level of risk. "At the end of the day, it is the customer that pays," says Hewgill.