Detecting fraudulent credit card transactions with artificial intelligence

Older Australians will doubtless remember physically swiping their credit cards in decades gone by, before credit cards with chips were introduced. Those were the good old days for credit card fraudsters. The technology to duplicate the simple magnetic stripe was inexpensive and easy to obtain, and large purchases required signatures which were easy to forge.

The introduction of the chip-and-PIN card a decade ago changed that. Unlike the magnetic stripe, the chips don't hold static data. Instead, they combine with the POS unit to generate a unique code for each transaction that is sent to the credit card issuer who verifies that this number can be generated from the correct card. These chip-and-PIN cards are much more difficult to skim.

The introduction of chip-and-PIN made it much more difficult to make counterfeit cards

There are, however, other ways in which credit cards can be fraudulently used. Lost or stolen credit cards can be used for payments with relative ease, especially with the introduction of contactless payment. In 2018, lost/stolen credit cards accounted for an all-time high of 10% of all credit card fraud. However, even this is small compared to CNP (card not present) fraud.

The rise of online shopping, with an average growth of 17.3% for the period 2015-2020 has led to a significant increase in CNP transactions, which are transactions where the credit card used for payment is not physically presented to the seller. CNP transactions are especially susceptible to fraud as the only information required is the credit card number and the CVV, which negates many of the security features included in modern credit cards. These details can be illegally obtained in many ways, such as through skimming, hacking, and phishing. Fraud resulting from CNP transactions accounts for more than 80% of all credit card fraud and grew quickly from $300m to $489m in the four years between 2014 and 2018.

Fraud resulting from CNP transactions accounts for more than 80% of all credit card fraud and grew quickly from $300m to $489m in the four years between 2014 and 2018.

Industry initiatives such as the CNP Fraud Mitigation Framework and greater awareness of credit card fraud have reduced the fraud value in recent years but this industry is still worth half a billion dollars annually in Australia. Significant amounts of time and effort go into identifying fraudulent transactions, with some of the onus falling on the merchants and some on the credit card issuer.

Detection of credit card fraud is no easy task, as evidenced by the continued prevalence of these kinds of transactions despite a decade of work to eliminate them. Fraudulent transactions account for only 3-4 in every 10,000, making identifying them akin to finding a needle in a haystack. This difficulty has led to most banks, credit card issuers, and large vendors adopting artificial intelligence (AI) techniques to augment human expertise. Companies such as AmEx, Paypal and VISA report significant decreases in fraud occurrence after the adoption of AI.

Companies such as AmEx, Paypal and VISA report significant decreases in fraud occurrence after the adoption of AI.

Historically, small retailers have had nothing of this sort and it has cost them heavily. But today you don't need to be a massive company to be able to leverage AI in reducing your risk of fraud. Deep Blue AI has built a suite of easily-accessible fraud detection solutions that democratise AI for payment transactions. If you're already spending money on fraud detection, using our solutions can help stretch that money further. And if you're not, why not?

Case Study

This case study uses data from transactions from European credit card holders over a two day period in September 2013. Transaction data was collected and anonymised by Worldline and the Machine Learning Group at Universete Libre de Bruxelles.

Of the 284,807 transactions recorded, only 492 were determined to be fraudulent - approximately 1.7 transactions per 1,000. One significant problem with classifying such rare transactions is that it is difficult for machine learning algorithms to spot patterns when there are such few examples of the transactions it should be flagging as fraudulent. Deep Blue AI's fraud detection solutions account for this with intelligent treatment of the few examples of fraud we do have.

When the item that is to be predicted is as rare as this, accuracy is a meaningless term. For example, we can achieve 99.83% accuracy by simply always predicting transactions to be genuine (since 99.83% of transactions are, in fact, genuine). But this would be useless in application since it would never pick out a fraudulent transaction.

Instead, we must look only at the examples we are interested in - the fraudulent transactions. For a model to perform well, it is necessary for it to maximise two metrics:

• The percentage of fraudulent transactions it flags - having a high number here indicates that the model has few false negatives.

• Accuracy for all transactions predicted to be fraudulent - having a high number here indicates that the model has few false positives.

Deep Blue AI's fraud detection algorithm was trained on 70% of the dataset described above, with the remaining 30% hidden from the model. On this hidden dataset, of all transactions the model flagged as fraudulent, 80% of those predictions were accurate. Of all fraudulent transactions in the dataset, 78% were found. The accuracy, although not the appropriate metric as discussed above, is 99.93%.

The picture below indicates the utility of such a model. The vast majority of transactions are genuine and the model only misclassifies 0.038% of these as fraudulent. Importantly, 114 of the 142 fraudulent transactions are picked up instantly.

Model performance on credit card fraud dataset

114 of the 142 fraudulent transactions are picked up instantly. This allows you to make additional checks such as identity or PIN at the point of sale without impacting the experience of genuine customers.

Deep Blue AI's solutions can be accessed via APIs or containerised, so they can be run on demand. For online transactions, this could be as soon as the customer checks out their basket, or when credit card details are entered. This allows you to make additional checks such as identity or PIN on those transactions predicted to be fraudulent and reduce the occurrence of fraud by 80.3% with no impact on the experience of genuine customers.

To see a demo of our fraud detection solutions, or to discuss your particular situation, contact us at contact@deep-blue.ai .