Document Fraud Detection via Landmark Noise Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting fraud in documents face challenges due to noise sources and limited data availability, which can lead to false rejections of genuine documents or acceptance of fraudulent ones, especially when there are only a few samples of a particular document type available.
Innovation Solution
A method is developed to train a machine learning model using 'original reference landmark locations' and 'modified' landmark locations synthesized through registration and noise modeling, allowing the model to handle limited data and noise variations, thereby enhancing fraud detection reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If machine learning is used to detect fraud using geometric measurements, then fraud detection capability is improved, but reliability deteriorates due to noise sources causing false rejections
Solution Approach 1:
The patent changes the parameter space from raw geometric measurements to transformed feature representations through registration and noise modeling. By transforming landmark locations through registration to a reference frame and incorporating noise models, the system creates a more robust feature space that accounts for variations in genuine documents, thereby reducing false rejections while maintaining fraud detection capability.
Solution Approach 2:
The patent introduces an intermediary processing layer between raw image data and fraud detection decisions. This intermediary consists of the registration process and noise modeling, which transforms and cleans the geometric measurements before they are used for fraud detection, filtering out noise that would cause false rejections.
2Quantity of substance
If machine learning model is trained with limited genuine document data, then data availability constraint is satisfied, but detection reliability deteriorates due to insufficient training data
Solution Approach 1:
The patent creates synthetic copies of genuine document features through the registration process. By transforming landmark locations from multiple genuine documents into a common reference frame, the system generates a virtual expanded dataset that maintains the statistical properties of genuine documents while increasing the effective training data volume, thereby improving detection reliability.
Solution Approach 2:
The patent performs preliminary registration and noise modeling on the limited genuine document data before training the machine learning model. This preliminary processing creates transformed feature representations that capture the essential characteristics of genuine documents, allowing the model to learn from the limited data more effectively and generalize better to unseen documents.
3Measurement precision
If geometric measurements are used to detect fraud, then precision of measurement is improved, but device complexity increases due to registration and noise modeling requirements
Solution Approach 1:
The patent segments the complex fraud detection task into distinct modular components: landmark detection, registration transformation, noise modeling, and fraud classification. Each component handles a specific aspect of the problem independently, making the overall system more manageable and maintainable while preserving measurement precision through specialized processing at each stage.
Data Source
AI summary
Described are methods and systems for detecting fraud in documents. First images of a first set of genuine documents and second images of a second set of genuine documents are obtained. A printed feature, spacings between printed features in the first images, and positions of printed features in the second images are selected. Selected features, spacings and positions are annotated to obtain original landmark locations for each printed feature, spacing and position. Annotated features, spacings and positions are transformed to obtain transformed features, transformed spacings and transformed positions. The transformed features, spacings and positions are combined with a noise model to generate modified features, modified spacings and modified positions. Each modified feature, modified spacing and modified position comprises annotations indicating modified landmark locations. Input data for a machine learning model is generated using original landmark locations and modified landmark locations. The machine learning model is trained using the input data.