Fraud detection is a challenging problem to solve and on the top to do list of every major bank. They can lead to huge revenue losses, bad customer experiences, loss of reputation and have a negative impact on growth.
In 2015, fraud losses incurred by banks worldwide reached $16 billion By 2020, card fraud worldwide is expected to increase to $36 billion.
With more customers moving from traditional banking methods to online banking, the number of fraud cases have been on the rise.
One of the main reasons for the increased number of online fraud, is the fact that most banks are still using traditional rule based methods of data analysis for detecting fraud. While rule based systems can be successful in detecting know patterns in fraud, they can falter in detecting new fraud patterns. Since online fraudsters are using sophisticated technology and creative methods, it is time for banks to up their game.
Here is an info graphic which shows the number of fraud cases from 2009-2019.
Number of bank fraud cases across India between 2009 and 2019.
Image source - https://bit.ly/2V9p9wa
Machine learning vs rule based systems
Rule-based methods of detecting fraud are complex, time consuming, require more human intervention, not real time and have a low fraud detection rate of lesser than 40%. Also, such systems can wrongly classify genuine customer transaction as fraudulent. Also traditional systems wont be able to adapt quick enough to catch relatively new methods of fraud.
Since most of these traditional systems have a human at the end, when the number of fraudulent transactions increases beyond a threshold, a bank might not have enough manpower to handle them. This could also lead to higher costs in the longer run.
AI Based fraud detection systems are more accurate at detecting fraud with an accuracy of more than 95%. They can also adapt quickly to detect new attempts at fraud in real time.
Why should machine learning be used in fraud detection?
Efficiency- The average number of transactions handled by banks per second is increasing year by year. Once a machine learning model has been built, these models can handle huge number of transactions in real time. They are also quick at adapting and finding new methods of fraud in short time.
Lower ownership costs- With cloud computing becoming more affordable, your cost of running machine learning models for fraud detection can be kept low. It can be scaled or reduced as per the transaction volume.
Greater adaptability- Banks that have offices in multiple geographies might need to combat different types of frauds in each region. Since AI algorithms are trained on data generated at each of these regions, they can adapt well to the local business needs.
How machine learning can help in detecting fraud?
Banks generate a huge amount of data on a daily basis. This data is used to train machine learning models which in turn helps in flagging a given transaction as fraudulent or not.
Fraud detection in machine learning comes under a topic known as anomaly detection (also known as outlier detection). This involves taking all the data and creating a graph. There may be a few transactions that might fall out of a given range. These are called anomalies or outliers.
For example, a given customer may be doing his ATM and online transactions from within a given geographical radius. There are services available which can map IP addresses to a physical address. If a transaction happens suddenly from a new IP address far away, this may be classified as a outlier and the transaction may not be allowed to happen.
Here is an image which illustrates this. Transaction IDs, source IP, start time, destination port, number of bytes are used to classify transactions as an attack or not.
Popular methods for Anomaly detection-
Support vector machines.
Clustering based anomaly detection- K means clustering.
Challenges faced while using machine learning to detect bank fraud
Lesser data- This is also known as the cold start issue. When there is not enough data, the prediction accuracy of the models built by the machine learning algorithms may be low. So smaller banks might have trouble in implementing such methods.
Lack of clarity- Sometimes it might be difficult to reason out why certain transactions were classified as fraudulent. Reverse engineering such cases might not be possible in certain scenarios.
Not knowing which algorithm to choose- The same machine learning algorithm might have different levels of success rates for different banks. Knowing which algorithm to use might be a challenge. It is recommended to work with a experienced machine learning solution provider in such cases.
Visit Skyl.ai to learn more about different machine learning and artificial intelligence projects.