Introduction
Veryfi Artifact Analysis provides sophisticated document forensics that work alongside our core data extraction API as part of the comprehensive Veryfi Fraud Detection Suite.
Artifact Analysis employs advanced computer vision and machine learning algorithms to detect three primary types of document manipulation:
Handwritten Characters: Manual alterations to key fields
Digital Tampering: Software-based document editing
AI-Generated Documents: Detection of artificially created receipts and invoices
The system returns both flags and clear color-coded results (Green, Yellow, Red) based on customizable thresholds, making it easy to integrate.
Handwritten Characters Detection
What is Handwritten Detection?
The system identifies handwritten content and provides a list of the exact areas where handwriting occurred using dot notation. The flag indicates the presence of handwriting - not necessarily fraud - giving you the flexibility to handle this information based on your specific use case and data patterns.
How does Handwritten Detection work?
Our machine learning algorithms analyze extracted field regions using trained models that recognize different handwriting styles and characters. The system examines specific fields for any traces of handwritten content and returns results in the handwritten_fields array within the document metadata.
Users can then apply custom business logic to determine appropriate handling based on their industry and document expectations.
What fields are examined for handwriting?
The default monitored fields include:
totalsubtotaldateline_items.totalline_items.pricetax_breakdown.taxtax_breakdown.tax_base
Note: Field configuration is customizable based on your specific requirements.
What are examples of handwriting scenarios and recommended handling?
Legitimate Handwriting Scenarios:
Restaurant receipts: Handwritten tips, totals
Service invoices: Handwritten descriptions, labor hours, or pricing
EU taxi receipts: Predominantly handwritten fare calculations and routes
Market vendors: Handwritten quantity and pricing information
Recommendation: Configure as low-risk or normal processing
Potentially Fraudulent Handwriting Scenarios:
Store receipts with handwritten modifications: Retail receipts typically use printed values
Altered line items: Handwritten product traces on product, quantity or price on printed receipts
Date alterations: Modified dates to fit reporting periods or campaign requirements
Amount modifications: Changed totals, subtotals, or quantities on printed receipts
Recommendation: Flag for review or additional verification
Possible Use Case-Specific Handling Examples:
CPG Loyalty Programs:
Store Receipt + Handwriting Flag → High Risk Review
Restaurant Receipt + Handwriting Flag → Standard Processing
Market Vendor + Handwriting Flag → Standard Processing
Expense Management:
Corporate Travel (Taxi) + Handwriting → Check Regional Patterns
Restaurant + Handwritten Tip → Standard Processing
Retail Store + Handwritten Total → Flag for Review
How is handwriting detected in the API response?
When handwriting is detected, the fraud analysis returns both the fraud color and a list of fields that are hadnwritten
{
"meta": {
"fraud": {
"color": "red",
"types": ["handwritten_characters"]
},
"handwritten_fields": [
"total",
"line_items.3total"
]
}
}Can handwriting detection be configured for different use cases?
Yes, handwriting detection offers extensive flexibility to accommodate diverse business requirements and data patterns. The system detects handwriting traces universally, but handling can be customized based on:
Industry-Specific Configurations:
CPG Campaigns: Configure strict handling for store receipts (handwriting = review)
Expense Management: Differentiate between restaurant (permissive) and retail (strict) handling
Configuration Options:
Field-level sensitivity: Customize which fields trigger different responses
Vendor category rules: Apply different logic based on merchant type
Geographic adjustments: Account for regional receipt practices
Threshold customization: Adjust fraud color assignment based on handwriting patterns
Business Rules Engine Integration: Create sophisticated workflows combining handwriting detection with other signals:
IF handwritten_fields contains "total" AND vendor_category = "retail_store" AND country_code = "US" THEN flag_color = "red" IF handwritten_fields contains "total" AND vendor_category = "restaurant" THEN flag_color = "green"
Learn more about Business Rules Engine
This flexibility enables organizations to fine-tune detection sensitivity according to their specific risk tolerance, industry requirements, and operational data patterns, ensuring the feature adds value rather than creating unnecessary friction.
Digital Tampering Detection
What is Digital Tampering Detection?
This feature visually detects instances where documents have been digitally altered or manipulated using software like Photoshop or similar tools. It identifies when fraudsters modify numbers, text in receipt images, including copy-paste manipulations such as duplicating digits to change amounts.
How does Digital Tampering Detection work?
The system analyzes pixel-level inconsistencies and pattern anomalies in photo submissions where visual alterations are present. Our computer vision algorithms are trained to detect the subtle artifacts left behind by digital editing software.
What types of digital tampering can be detected?
What fields are analyzed for digital tampering?
Configurable default fields include:
totalsubtotaldateline_items.totalline_items.pricetax_breakdown.taxtax_breakdown.tax_base
What file formats support digital tampering detection?
Supported:
Photo submissions (JPG, PNG, HEIC, etc.)
Image files where visual alterations are present
Limitations:
Requires visual evidence of manipulation
Works best on high-resolution images
Detection accuracy varies based on editing sophistication
How is digital tampering detected in the API response?
When digital tampering is detected, the system returns:
{
"meta": {
"fraud": {
"color": "red",
"types": ["digital_tampering"]
}
}
}AI-Generated Document Detection
What is AI-Generated Document Detection?
This feature identifies documents created by artificial intelligence tools such as ChatGPT, DALL-E, Midjourney, or other generative AI systems. As AI-generated content becomes more sophisticated, this detection capability helps prevent fraud attempts involving artificially created receipts that never existed in reality but appear legitimate.
How does AI-Generated Document Detection work?
The system employs a dual detection approach that analyzes both visual patterns and metadata signatures typical of AI-generated content. Our algorithms are trained to recognize the subtle artifacts, inconsistencies, and patterns that distinguish AI-created documents from authentic photographs of physical receipts.
Detection Methods:
Visual Analysis: Identifies AI rendering patterns, lighting inconsistencies, and geometric anomalies
Content Pattern Recognition: Detects unrealistic text layouts, font combinations, and formatting typical of AI generation
Metadata Examination: Analyzes creation signatures and processing patterns associated with AI tools
What types of AI-generated fraud can be detected?
Common AI Generation Scenarios:
Expense Management:
AI-created receipts for non-existent business meals
Generated hotel or travel receipts with realistic merchant details
Fake service provider invoices with convincing layouts
CPG Loyalty Campaigns:
AI-generated store receipts featuring qualifying products
Artificial promotional receipts with campaign-specific items
Generated receipts from target retailers or locations
Fake bulk purchase documentation for bonus redemptions
How sophisticated are current AI-generated documents?
Current AI Capabilities:
High-quality visual rendering with realistic layouts
Convincing merchant names and product descriptions
Proper formatting and mathematical calculations
Realistic lighting and paper textures
Detection Advantages:
AI-generated content often contains subtle inconsistencies in shadows, reflections, or text rendering
Mathematical relationships may be perfect in ways real receipts rarely are
Font and spacing patterns may not match authentic printing methods
Color gradients and paper textures may show artificial generation artifacts
How is AI-generated content detected in the API response?
When AI-generated patterns are detected
{
"meta": {
"fraud": {
"color": "red",
"types": ["generated_document"]
}
}
}Can AI-Generated Detection be configured?
Yes, the fraud decision color for this feature is fully configurable. Organizations can adjust sensitivity based on:
Industry risk tolerance: High-value expense programs may use stricter thresholds
Campaign value: Premium loyalty programs may require maximum sensitivity
Volume considerations: High-submission campaigns may balance detection with processing efficiency
Integration with other signals: Combine with handwriting and tampering detection for comprehensive protection
For detailed technical information and implementation guidance, see: Detecting AI-Generated Documents with Veryfi
How do I configure Artifact Analysis thresholds?
Can I combine Artifact Analysis with other fraud signals?
Yes! Artifact Analysis works as part of the comprehensive Veryfi Fraud Detection Suite, which includes:
Bad Actors Detection:
High velocity submissions
Multiple devices/profiles
Emulated device detection
Classification Analysis:
Non-document identification
Screenshot detection
AI-generated content identification
Uniqueness Verification:
Duplicate document detection
Similar document analysis
Use Case Applications
Expense Management Implementation
Primary Benefits:
Prevent manually altered expense claims
Detect sophisticated digital manipulation
Maintain audit compliance standards
Implementation Strategy:
Start with handwriting detection for high-value claims
Add digital tampering for photo submissions
Include AI-generated detection for all submission types
Configure thresholds based on risk tolerance
Common Integration Patterns:
High Value Claims → All three detection types
Standard Submissions → Handwriting + Digital Tampering
Mixed Submissions → AI-Generated + Digital Tampering
CPG Loyalty Campaign Protection
Primary Benefits:
Ensure campaign integrity across all submission types
Prevent fraudulent redemptions and promotional abuse
Protect marketing budgets from manipulation
Maintain consumer trust in promotional programs
Campaign-Specific Configurations:
High-Value Promotions: Maximum sensitivity across all detection types
Volume Campaigns: Balanced detection to prevent false positives
ROI Protection:
Prevent budget drain from fraudulent submissions
Maintain fair competition among legitimate participants
Reduce manual review overhead through automated detection
Best Practices & Implementation Guide
Getting Started with Artifact Analysis
Phase 1: Assessment
Contact your Veryfi account manager or [email protected] to enable Fraud Suite
Configure initial field lists based on your use case using Veryfi Business Rules
Set conservative thresholds for initial deployment using Veryfi Business Rules
Implement logging for fraud pattern analysis
Phase 2: Testing
Run parallel analysis on historical submissions
Analyze false positive/ true positive/false negative rates
Create your internal metrics that
Adjust thresholds based on results
Phase 3: Production Deployment
Implement automated workflows for each fraud color
Set up monitoring and alerting
Create manual review processes for yellow/red flags
Workflow Integration Recommendations
Automated Actions by Fraud Color:
Green Submissions:
Process normally through the standard workflow
No additional review required
Yellow Submissions:
Flag for expedited manual review
Route to the specialized fraud investigation team
Implement additional verification steps
Red Submissions:
Block automatic processing
Require supervisor approval
Trigger comprehensive audit procedures
Monitoring & Optimization
Key Metrics to Track:
Detection accuracy rates by fraud type
False positive/negative rates
Manual review volume changes
Financial impact prevented
Continuous Improvement:
Regular threshold adjustment based on patterns
Seasonal configuration updates for campaigns
Integration of new detection capabilities
Training updates for review teams
Limitations & Considerations
Current Technical Limitations
Handwritten Characters:
Limited to configured fields (customizable)
Requires clear handwriting visibility
May not detect very subtle alterations
Digital Tampering:
Requires visual evidence in images
Sophisticated editing may reduce detection accuracy
Cannot analyze compressed or heavily processed images effectively
AI-Generated Documents:
Detection accuracy improves as AI generation patterns are learned
Extremely sophisticated AI models may occasionally evade detection
Performance varies based on AI tool sophistication and generation quality
File Format Support
Fully Supported:
High-resolution photo submissions (JPG, PNG, HEIC)
Standard document images for all detection types
Scanned document images
Detection-Specific Support:
Handwriting Detection: Works on all image formats with clear field visibility
Digital Tampering: Requires uncompressed images with visible manipulation artifacts
AI-Generated Detection: Effective across all image formats and some document types
Limited Support:
Heavily compressed images
Low-resolution submissions
Screenshots (detected by other fraud signals but limited manipulation analysis)
Documentation & Resources
Need help? Reach out to [email protected]