Fraud Detection: Fighting Financial Crime with Machine Learning

remember this scene from the office with kevin and the credit card no not that one this one when jim and pam go on a honeymoon to puerto rico kevin gets comfortable in jim's office and receives a call from jim's bank well mr halpert you're obviously not in san juan puerto rico jim's card gets blocked and it results in a few gags when kevin tries to reach jim by phone and learns that he fortunately doesn't know it was kevin's fault this episode came out in 2009 more than 10 years ago back then there was an important travel rule let your bank know when and where you're going so they don't block your cards for suspicious activity around 2015 many large banks stopped asking their clients to report their travel plans and those that still ask you to do that won't block your card automatically but will send you a message for purchase approval that's because security has come a long way when it comes to preventing financial fraud banks no longer activate code red when they see you buying coffee at an airport in singapore but how are fraud detection and prevention technologies so nuanced they can distinguish between customers adventures and fraudsters schemes fraud detection fighting financial crime with machine learning everyone is exposed to financial fraud if you're selling or buying something online providing financial services or simply processing tons of payments you face fraud risks every day for businesses scams are especially scary because you're not only losing money but also customers who may no longer trust you so detecting and preventing fraud is essential there are two approaches to catching fraud the more common one is using rules the more effective one applies machine learning let's start with the first one to show how rule-based systems work let's look at uber one common type of fraud the company encounters is when riders are paid via stolen credit cards behind every fraud there's a structure a pattern that becomes visible if you look close enough so to detect fraudulent rides uber looks for similarities between accounts that committed fraud using analytics for example they found that many fraudulent accounts were registered in asia while making the trips in the biggest cities around the world like london or new york these accounts were also created within mere seconds from one another had gibberish in their email addresses and often use the same credit card the geolocation of account creation revealed even more logic behind criminals minds see how accounts respond in the grid-like formation and some were even created in water all these pretty solid signals enabled uber to locate potential fraudsters and stop them using rules rules are written using if this then that statements for example if the account was created in china and is requesting for a ride in new york then don't dispatch uber's rules engine takes thousands of signals into account and there are many complex rules that help stop fraudsters the problem is that fraudsters are smart too and when they catch on to the fact that their accounts created in china are blocked they can switch to a different ip one in australia or norway or wherever so it's like the security system and fraudsters play an endless game of one upmanship trying to get around one another and learning from each other in the process and that takes a lot of time and effort rules are written manually by analysts finding a new signal and creating a rule for it can take days or weeks so some fraud always slips in or goes unnoticed modern security systems have a solution for that machine learning [Music] machine learning is the accessible way to implement artificial intelligence in real life we have a whole playlist of videos explaining machine learning and ai concepts so be sure to check some out after this video one thing you want to keep in mind is that machine learning based systems are extremely good at finding patterns that humans can't you see where this is going right machine learning knows very nuanced distinctions between normal and fraudulent activities and finds way more signals than human analysts do here's how it does that step 1 understanding what is normal to be honest frauds don't happen that often and you don't want to be alerted about suspicious behavior all the time basically you need to minimize false positives like people who simply want to use their credit card on vacation so a model has to know what behavior is normal to identify what deviates from that normal and what signals point to fraudulent activity so it's first trained on historical data historical data are past transactions both fraudulent and not each transaction stores all different attributes aka features about it the exact time of transaction purchase amount credit card number shipping address sometimes several attributes are combined to create a feature like the transaction date and the date of the customer's past transaction basically everything of course not all of this information matters when looking for fraud so a data scientist must go through all those features and mark the ones with the largest predictive power normally the most useful features revolve around transaction recency time passed since last purchase frequency number of transactions and monetary values dollar amount of transactions if a customer made several large transactions within a day there's an excellent chance their account was hacked at the same time if someone made a purchase in beijing and then in 11 hours another purchase in delhi it's plausible that the same person physically traveled between those cities and was able to complete valid transactions so transactions made in different locations don't always indicate fraud you'd have to combine attributes into another feature a combination of distance between these locations and times of transactions after all important features are picked the model is trained to learn the patterns in the historic data and predict what transactions were valid or not step 2 finding anomalies one of the most common targets of financial fraud today are e-commerce merchants because opening new accounts or accessing customer accounts is not that hard online stores are prone to scam here fraud is also solved by distinguishing normal customer behavior from abnormal this is called behavioral analytics in behavioral analytics all customers are segmented into groups so the system can easily identify if one customer's behavior fits the behavior of the group here's what happens at the exact time of the transaction first a customer makes a purchase then features of one transaction are compared to the features of other buyers basically checking if this is the typical behavior for people in this segment for example fraudsters usually make transactions during non-working hours but your late night shoppers might as well and frequent travelers don't pose a risk if all other features of their transactions make it valid finally after considering the predictive power of all features a machine learning model calculates a fraud score an algorithm doesn't make decisions based on a yes or no answer it actually returns the probability of the fraud as a customer's purchase activity is rarely squeaky clean the system calculates how likely it is that this exact purchase is fraud when the fraud score is low the purchase is approved when high the system flags the transaction as suspicious to be reviewed by the store or the customer step 3 eliminating mistakes many advanced fraud detection systems are effective in 99 of cases the remaining 1 are errors false positives and false negatives as we already mentioned false positives occur when the model detects fraud where there's none false negatives are fraudulent transactions that slip through but since there are many more valid transactions than fraudulent ones increasing the effectiveness of your model will also boost the number of false positives which is a hit to customer satisfaction a card declined can cost you a customer this problem is handled by using a machine learning method called deep neural networks neural networks use a learning mechanism inspired by the one in our brains they allow us to find non-linear relations between a massive amount of data points without manually selecting features these relations can be so nuanced that we would fail to understand the logic behind the final decision this complex processing allows neural networks to learn from much more pieces of information than traditional ml models do and find even more fraud signals making fraud detection systems almost error proof but why does fraud still happen it seems that with all this technology guarding us we should be pretty safe from financial fraud yet it keeps happening how first technology doesn't just serve the good guys fraudsters benefit as well machine learning on-demand computing power and data analytics helps scammers commit fraud using cheaper and more sophisticated methods and if your security system doesn't keep up you won't be able to protect yourself scammers also adapt quickly and they're especially good at finding loopholes in changing conditions in this era of the pandemic when people are more vulnerable and put trust in third parties fraudsters come up with new ways to target them in 2020 alone the ftc has logged more than 324 000 complaints related to covid19 69 of which involved identity theft insurance and healthcare organizations suffered too with fraudsters taking advantage of the situation to exaggerate claims there's also an ever-growing strain on businesses to make customer experiences frictionless while keeping their systems as secure as possible finding such balance often involves more advanced technology like deep learning which is honestly not available to many retailers insurers or banks so they play catch up with criminals who often stay one step ahead but businesses have the upper hand they just don't always use it they already collect and store vast amounts of data about transactions and users but haven't yet learned to organize and analyze it to its full potential the potential that only ai can uncover [Music] you

Channel: AltexSoft

Views: 59,105

Keywords: Fraud detection, scam, machine learning, rule-based fraud detection, Neural networks, false positives, false negatives

Id: QFyM3w95fXI

Length: 12min 0sec (720 seconds)

Published: Tue Jul 27 2021