Mitigating Bank Scams with Machine Learning
Fighting Fire with Fire
The latest Annual UK Finance fraud report highlights that the value of Authorised Push Payment Scams decreased by 17% last year when compared to 2021. Despite this still amounting to a whopping £485.2 million lost, the results do demonstrate progress in Banks’ fraud detection strategies and customer education.
Looking deeper into the types of scams recorded, we see that impersonation scams saw the largest decline of 20% in value. These usually involve the fraudster impersonating a Bank, the Police or another trusted authority and typically result in larger value losses, such is the sophistication of the fraud.
On the other hand, the report also shows that Purchase Scam fraud volume and value have actually increased by 17% and 4% respectively. A purchase scam can range from buying a new smartphone via an advert seen on Facebook marketplace, which never arrives or buying a pair of designer trainers from an Instagram sponsored post, which turn out to be counterfeit.
These types of scams tend to be lower in value but very high in volume, as is demonstrated by the number of fraud cases reported. The continued rise in mobile banking, and the fact that mobile is the preferred channel of choice for customers’ social media, shopping, gambling and emails, likely goes a long way towards explaining this. Essentially it is very easy to see something online and then jump into your mobile banking app and make the purchase. As such, these scams are typically harder to detect than other type of scams, since banks’ fraud detection tools will consider them a perfectly normal purchase.
LexisNexis® Digital Identity Network®
Machine Learning Payment Model
Purchase scam Vs Impersonation scam behaviours
Figure 1:
Never Before Seen
70% of impersonation scams were made to brand-new beneficiaries not seen before in the Digital Identity Network, whereas less than 50% of purchase scams were made to new beneficiaries in the network.
Figure 2:
Active Call in Progress
In over a third (33%) of impersonation scams, the customer was on an active call (i.e. a phone call was in progress on the device) when making the payment, whereas just 6.5% of purchase scams were on an active call.
Figure 3:
Working 9-5
Impersonation scams were more likely to occur during typical 9-5 working hours and the customer was seen to be logging in twice as much in the preceding hour – perhaps to verify the scammers’ claims – when compared to purchase scams.
Figure 4:
Every Second Counts
The average time taken to make the payment was 90 seconds longer with an impersonation scam, compared to a purchase scam.
The Results
With this stark difference in behaviours between the two scam types, the data science team at LexisNexis Risk Solutions were able to develop and assess performance over multiple scam models. Results showed that the machine learning model performed best when purchase scams were excluded, compared to training the same model on all fraud types. By excluding purchase scams, we were able to increase detection on impersonation scams by 8% and other scam types, such as remote access tool (RAT) scams – where the fraudster gains control of the victim’s device - by 15%. These two scam types alone were the most significant in value in terms of losses for the Bank.
Overall, the model achieved +£2.7m annualised incremental rise in the total value of scams prevented at 0.1% review rate – equivalent to 95 cases reviewed per day – compared to the existing incumbent model.
The annualised total value of the scams detected by the model at this alert level was £4.3m, at a detection rate of 34.8%, broken down by typology as follows:
- Impersonation Scams – 56% of total scam value detected
- Remote Access Scams – 64% of total scam value detected
- Investment Scams – 25% of total scam value detected
Summary
LexisNexis Risk Solutions machine learning custom scam models, combined with functionality such as Active Call Detection and Global beneficiary intelligence play a pivotal role in detecting scams. By creating customised models for our clients, we can tailor the performance with their operational capacity in mind, making sure that our clients get the maximum benefit from the model.
The Bank’s ability to accurately mark their scam losses into distinct scam categories has had a direct and tangible benefit. It allows for better separation of behaviours and increased performance in detecting high-value scams such as impersonation and remote access (RAT) scams.
Where there is a scam taking place there is also a money mule at the other end waiting to receive the fraudulent funds. Our data science team have developed machine learning mule propensity models utilising functionality such as Advanced Payment Screening – assessing incoming and outgoing payments –to detect the likelihood that a payment is being sent to a mule account.
As Banks start to adjust to new reimbursement regulations following the announcement by the Payment Systems Regulator (PSR) in June 2023, the focus on scam detection, scam customer education and mule strategies is as important as ever before. The changes will no doubt be significant, and the impact may be felt harder on some banks, where there are deficiencies in scam and mule controls.
Joey Bajela, Lead Engagement Manager, LexisNexis® Risk Solutions
