	US 12,354,411 B1
	Apparatus and method for assigning financial transaction document images to respective models for fraud detection based on document image characteristics
Atul Mittal, Karnataka (IN); Sonu Agarwal, Karnataka (IN); Joe Manjiyil, Karnataka (IN); Neeta Pande, Karnataka (IN); and Sandeep Ramesh, Plano, TX (US)
Assigned to Morgan Stanley Services Group Inc., New York, NY (US)
Filed by Morgan Stanley Services Group Inc., New York, NY (US)
Filed on Oct. 15, 2024, as Appl. No. 18/915,921.
Int. Cl. G06V 40/30 (2022.01); G06V 30/22 (2022.01); G06V 30/41 (2022.01); G06V 30/42 (2022.01)

CPC G06V 40/33 (2022.01) [G06V 30/22 (2022.01); G06V 30/41 (2022.01); G06V 30/42 (2022.01)]

20 Claims

1. A method, comprising:

obtaining, at a processing apparatus, a plurality of document images;

performing, at the processing apparatus, a material alteration detection process and a separate signature forgery detection process on the obtained plurality of document images, wherein

the material alteration detection process comprises:

automatically sampling, at the processing apparatus, a subset of the obtained plurality of document images based at least in part on account information indicated on the obtained plurality of document images;

performing, at the processing apparatus, image pre-processing on the sampled subset of the obtained plurality of document images;

determining, at the processing apparatus, a document type for each image of the sampled subset of the obtained plurality of document images, the document type indicating one of at least a handwritten document type and a printed document type;

upon determining the handwritten document type for one or more document images of the sampled subset of the obtained plurality of document images, analyzing, at the processing apparatus, the one or more document images using a machine learning (ML) algorithm trained to detect material alterations on handwritten documents; and

outputting, at the processing apparatus, a fraud probability representation for each analyzed document image, and

the signature forgery detection process comprises:

performing, at the processing apparatus, image pre-processing on the obtained plurality of document images;

obtaining, at the processing apparatus, one or more past signatures corresponding to each of the obtained plurality of document images;

performing, at the processing apparatus, signature image pre-processing on each obtained past signature;

authenticating, at the processing apparatus, each signature included in the obtained plurality of document images using the obtained one or more past signatures and a ML algorithm trained to match signatures;

outputting, at the processing apparatus, a similarity measure for each authenticated signature from the ML algorithm trained to match signatures;

adjusting, at the processing apparatus, a threshold associated with the outputted similarity measure based on additional account information associated with each said authenticated signature;

comparing, at the processing apparatus, the outputted similarity measure to the adjusted threshold; and

outputting, at the processing apparatus, an alert upon determining that the outputted similarity measure fails to meet the adjusted threshold,

wherein the ML algorithm trained to detect material alterations on handwritten documents is trained by under-sampling “non-fraud” documents.