Boosting fraud detection accuracy in banking
Context & Objectives
A bank in Belgium is dealing with the growing problem of document forging and mortgage application documents fraud. This results in giving people new loans or increasing the size of their loans when they won't (or can't) pay them back.
The banks' current process to detect document fraud is for individual agencies to check each document manually. Once an agent flags a document as suspicious, they refer it to a team of experts for review. Unfortunately, this slow and tedious process can easily result in human error, such as passing fraudulent documents as genuine.
This bank tasked Agilytic with providing a proof of concept for an automated fraud detection system. This system must achieve two main objectives:
Save employees time by removing manual checks of every document.
Improve the accuracy of detecting fraud.
Approach
We took three approaches to develop an automated fraud detection system:
The first two are traditional approaches based on optical character recognition (OCR) to detect alterations in text documents. These compare value and format inconsistencies in documents, such as spaces between letters. OCR does not require large amounts of training data so long as documents follow a specified template that the OCR system can analyze.
The third is an innovative approach involving steganographic fraud detection to detect manipulations on image documents. This graphical inspection technique can detect fraud in even the trickiest cases, going beyond what our eyes can see and extending down to the pixel level of these image documents. We tested our algorithm for this approach on 2,300 pay slips, and it correctly detected 70% of the forged documents.
Combining these three approaches significantly improves the system's ability to detect fraud accurately. While OCR offers the advantage of implementing a simple solution incredibly quickly, steganography allows us to reach high levels of accuracy in cases where the human eye may have easily overlooked the presence of fraud.
Results
We designed the system to make it quick and easy for agencies to check a document for fraud. The system either correctly detects the fraud outright or provides recommendations around whether the document contains fraudulent elements that the agent can then review themselves.
Using the information provided by the system as a guide, agencies can reach a conclusion about the viability of a document a lot faster, resulting in a much more streamlined process.
A key point to remember when automating processes that previously involved highly manual tasks is that you may miss the point of 100% automation. However, automating even a fraction of the tasks that are high-volume and error-prone can bring enormous benefits to an organization.
In this case, the automated fraud detection system could correctly detect around half of all submitted documents, significantly reducing the workload of agents at the bank.
The high-cost nature of detecting fraud for banks makes the automation of the process an incredibly worthwhile endeavor. In addition to saving time and improving the accuracy of detecting fraudulent documents, automated fraud detection systems can help banks manage risk more effectively by identifying and addressing fraudulent activity before it becomes a problem.
—
Written by Joleen Bothma
