Operationalizing AI & Self Learning in Check
Operationalizing AI in Check
Neural networks (NN), CAR, LAR and ICR have had a significant positive impact in the evolution of check processing into an omnichannel capture environment. If it wasn’t for the efficiencies of check recognition and image analysis, industry-wide adoption of image capture may have never happened because much of the labor savings was driven by amount and field processing automation.
Throughout the years, NNs have illustrated strengths and weaknesses. As an example, OrboGraph operationalized and deployed a 3rd party NN engine in 1999. At the time, our approach to courtesy amount recognition (CAR) handwriting recognition included algorithmic programming coupled with this NN system. For several years, this combination was the best in the market, performing at 60-75% read rate.
However, this NN became outdated as OrboGraph created higher performing solutions which utilized legal amount recognition (LAR). This foundation created a path to virtually 100% recognition.
With the introduction of OrbNet AI technologies with deep learning models, the recognition process has evolved with a path to virtually 100% recognition and fraud prevention rates can now exceed 95% with OrbNet Forensic AI.
Operationalizing Self Learning in Check Fraud
In the Exploration of Check Fraud in 2020 survey created by OrboGraph, it was confirmed that financial institutions have a growing problem when dealing with on-us and deposit fraud in check payments. Legacy fraud prevention technique involving filters, and even solutions using data analytics or machine learning, fail to detect counterfeit, forged, and altered items due to the nature of these types of fraud.
To help solve this problem, FI's must use image analysis along with a transaction analysis approach coupled with self learning capabilities.
OrbNet Forensic AI performs image analysis in several targeted ways:
- Check stock validation (CSV) for counterfeit detection
- Automated signature verification (ASV) for forgery identification
- CAR/LAR discrepancy and payee name verification (PNV) for alteration detection
In order for these image analysis technologies to be effectively deployed, self learning enables the system to adapt to the image and transactional characteristics of each account, storing these behaviors along with image snippets for real-time fraud detection capabilities. This self learning approach is based on thresholding rules, flagging major variations in check stock and signatures which may indicate a potential fraudulent payment. Small variations are automatically added to each account profile while larger variations require operator review. Once reviewed, feedback is provided back to the system to validate a “good” or “bad” payment. These account profiles become mature based on the learning history of the transactions. The result is a fraud detection process which identifies a high degree of fraudulent items with minimal suspect levels (False positives).
OrboGraph continues to invest in intelligent payment automation technologies for check processing including AI, deep learning and self learning within the OrboAnywhere suite. It is our quest to drive straight-through processing for the industry as well as protect financial institutions and their clients with best-in-breed image analysis.
