Veryfi Fraud Detection Suite
Veryfi Fraud Detection Suite runs alongside Veryfi document extraction. It flags documents that look risky before they move through your workflow, and returns a simple color-coded result (green, yellow, red) based on thresholds you control. Fraud results come back in the same API response as your extracted data, so there is no separate call to make.
This article covers the Receipts / Invoices API: https://api.veryfi.com/api/v8/partner/documents
For developer reference, see the Veryfi API docs: Fraud overview, Signals, Triggers, and Fraud types. And the Full API Schema >meta object
Why fraud detection matters
Receipt and invoice fraud is a growing problem for consumer packaged goods (CPG) brands, loyalty and rewards programs, and FinTech platforms. Catching it early helps you:
Protect assets: stop paying out rewards, rebates, or reimbursements on fake or altered documents.
Keep customer trust: honest users are not penalized for the behavior of a few bad actors.
Reduce cost: fewer manual reviews and fewer chargebacks or write-offs.
Support compliance: keep an auditable record of why a document was flagged.
Protect your brand: keep promotions and programs from being gamed.
Keep competition fair: promotions and contests reward real purchases.
Common patterns Fraud Suite helps catch include fake or edited receipts, AI-generated receipts, promotion stacking, high-volume submissions from one device, and the same receipt submitted more than once.
What you get
Layered protection: device checks and document checks work together.
Your thresholds: turn each signal on or off and set its sensitivity.
Easy integration: results come back in the same API response as your extracted data.
Ongoing improvement: new signals are added over time, and your review feedback helps tune detection.
How Fraud Suite works
Fraud Suite checks documents in two stages:
On the device. If you capture documents with Veryfi Lens, Lens runs device security checks, image quality checks, and source verification at the moment of capture and either prevents the user from submitting an illegitimate document OR passes the device data to Veryfi backend for further evaluation of device checks. Velocity tracking requires Lens.
In the cloud. Veryfi AI analyzes the document using computer vision, natural language processing, and behavioral pattern recognition to catch more sophisticated fraud.
Signals and the decision
Every check produces a result, and Fraud Suite turns those results into one decision for the document. The API returns both:
Signals. Each check (tampering, AI generation, duplicates, velocity, and so on) returns its own result. Most return a value and a score.
Decision. Fraud Suite is a logic layer on top of those signals. It reads them, applies your Fraud Config (which signals are on, their thresholds, and the color each one can force), and rolls everything up into one decision: a color, a score, and the list of fraud types that fired.
Signals (fields + scores) → Fraud Config (on/off, thresholds, colors) → Decision (color, score, types, details)
Choose what you consume
Different teams usually want different layers. You can use the decision, the raw signals, or both.
| Use the decision | Use the raw signals |
Good fit for | Routing documents to approve, review, or reject queues. Ops and review teams that want Veryfi's verdict. | Teams with their own risk engine or rules, or that want to combine Veryfi signals with their own data (user history, amounts, policy). |
What you read | meta.fraud.color, meta.fraud.types, meta.fraud.details | The Raw field column in the fraud type tables |
Where thresholds live | In Fraud Config. Changes take effect without code changes on your side. | In your code. Fraud Config thresholds do not change the raw signals response or scores. |
What to expect | One color for the whole document. It is not a per-signal verdict, and it does not explicitly say which signal drove it. But lists all fraud types triggered. | A field can show a high score while the type did not fire, because a type only fires after you set the threshold in Fraud Config and activate it. |
Using both. A common pattern is to route on color and store the raw scores so you can audit decisions and tune Fraud Config over time.
Recommended starting point. Use both device checks (Veryfi Lens) and document checks, and start with the default thresholds. Adjust once you have seen results on your own traffic in Fraud Stats. Not all fraud signals may be relevant to your use case; enable or disable signals to affect the color in Fraud Config. Note: raw results will be delivered regardless of this configuration. Configuration only controls the traffic color logic.
Before you build on raw signals
Not every check has a raw field. fraud history, emulated device, blocked device, suspicious content, data inconsistency are only returned as a type in
meta.fraud.types, with an explanation inmeta.fraud.details.If you want these, read types even if you ignore color.Not every field lives under
meta.fraud. Handwriting fields andmeta.pages[]are undermeta. But e.gis_duplicateis on the document level. Copy paths from the tables or API Schema exactly to build correct mapping.There are two pages arrays. LCD photo is on
meta.fraud.pages[]. Generated document and screenshot detail are onmeta.pages[].Missing and
nullmean different things per signal.meta.fraud.fraudulent_pdfalways has every key, andnullmeans that the check could not answer.meta.fraud.fraudulent_imageis missing entirely when the signal is off. A missing object is not a clean result.Warnings are a separate input.
meta.warningsnever affect color, score, or types, but you can use them in your own logic.
Color-coded decision
meta.fraud.score is the probability, from 0 to 1, that the document is fraudulent. The score sets the color:
Color | Meaning | Score |
🟢 Green | Low risk | 0.50 or below |
🟡 Yellow | Moderate risk. Review recommended. | Above 0.50, up to 0.75 |
🔴 Red | High risk. Likely fraudulent. | Above 0.75 |
meta.fraud.score
When only one signal fires, the document score equals that signal's score.
When multiple signals fire, the document score equals to the highest individual contributor's score.
Some signals can also force a color on their own, regardless of score. For example, definitive is_duplcate, emulated device and blocked device are immediately red.
Decision fields
Field | What it is |
meta.fraud.color | green, yellow, or red. Anything other than green should go to review or an automated workflow. |
meta.fraud.types | The fraud types that fired after your Fraud Config thresholds, sorted by importance |
meta.fraud.details | A human-readable description for each fired type, plus reasons for some. See What details looks like. |
meta.fraud.attribution | The detected types as one comma-separated string, ready to display to end users, repeats types in a user-friendly way |
meta.fraud.score | One score from 0 to 1 for the document. It sets the color. |
To find out which signal drove the color, read types and details. Color alone does not tell you.
Example
One signal fired: the digital background model returned a score of 0.98, which passed the account's threshold.
"meta": {
"fraud": {
"color": "red",
"score": 0.98,
"types": ["digital background", "screenshot"],
"attribution": "digital background, screenshot",
"digital_background": {"value": true, "score": 0.98},
"not_a_document": {"value": false, "score": 0.03},
"details": {
"digital_background": {
"description": "Digital-background artifacts were detected in the document image."
}
}
}
}
How to read it:
screenshotis the legacy name for the same check asdigital background, so this is one signal, not two.Because it is the only signal that fired, the document score equals the signal's score: 0.98.
0.98 is above 0.75, so color is red.
not_a_documentis false, so it did not fire and is not in types.
-> Decision consumers read color, score, types, and details.
-> Raw signal consumers read digital_background, not_a_document, and the other signal fields.
Naming changes in progress
Code against the current names. Expect the old names on some responses for now.
Current name | Old name | What to do |
digital background | Screenshot | Both types are returned during the rename. Treat them as the same check. screenshot will be deprecated. |
generated document | ai generated | Older documents from 2025 may still return ai generated. |
suspicious content | stopwords (Fraud Config only) | Fraud Config may still store the phrase list under stopwords. The API type is suspicious content. |
Fraud types
Fraud Suite groups its signals into three categories:
Device signals: catch bad actors at the source, prevent fraudulent or illegitimate submissions using Veryfi Lens.
Document and image signals: look for edits, AI generation, screens, and non-documents.
Data and content signals: compare what the document says against other documents, prior evidence, and itself.
Each signal can be turned on or off, and its sensitivity set, in Fraud Config.
Device signals
These depend on Veryfi Lens, except blocked device, which uses your own blocklist.
Type | Default color | Details key | Description | Raw field |
high velocity | Yellow | high_velocity | Unusually high submission velocity was detected from this device. | meta.fraud.submissions |
critical velocity | Red (off by default) | critical_velocity | Critically high submission velocity was detected from this device. | meta.fraud.submissions |
multiple profiles or devices | Configurable | multiple_profiles_or_devices | Multiple user profiles were detected on the same device. | meta.fraud.device_profiles |
fraud history | Raises the score | fraud_history | Prior fraud signals were detected for this device or user. | None |
emulated device | Red | emulated_device | The document was submitted from an emulated device. | None |
blocked device | Red | blocked_device | The submitting device is on the client's blocked-device list. | None |
All device signals return a description only in details.
Document and image signals
Type | Raw field | Description |
handwritten characters | meta.handwritten_fields (list) | Handwritten characters were detected on one or more extracted fields. |
handwriting tampering | meta.handwritten_tampered_fields (list) | Printed content on the document was overwritten by hand. |
digital tampering | meta.fraud.digital_tampering_fields (list) | Digital tampering was detected on the document image. |
not a document | meta.fraud.not_a_document {value, score} | The submitted image does not appear to be a valid document. |
LCD photo | meta.fraud.pages[].is_lcd {value, score} per page | The document image appears to be a photo of an LCD screen. |
digital background | meta.fraud.digital_background {value, score} | Digital-background artifacts were detected in the document image. |
generated document | meta.pages[].ai_generated {value, score} per page | The document appears to have been AI-generated. Has reasons |
fraudulent pdf | meta.fraud.fraudulent_pdf (component scores) | The PDF contains indicators of editing or manipulation. Has reasons |
fraudulent image | meta.fraud.fraudulent_image (component scores) | The document image contains indicators of editing or re-encoding. Has reasons |
Data and content signals
Type | Raw field | Description |
duplicate | is_duplicate | A previously submitted document matched this submission based on image or text similarity. |
similar documents | meta.duplicates (list) | This document is highly similar to other previously submitted documents. |
suspicious content | None | The document contains language or values commonly associated with mock, test, placeholder, or sample content. |
data inconsistency (beta) |
| Details are coming |
Every fired type returns a description in details. Types marked Has reasons also return a reasons list, and so does the legacy screenshot key.
A closer look at each signal
Device signals
High velocity and critical velocity.
Many submissions from one device_id. meta.fraud.submissions counts submissions over several time windows (minutes, day, weeks, month). Critical velocity uses the same field with higher thresholds, and is off by default. Both thresholds are configurable.
Multiple profiles or devices.
One account used on several devices, or one device used by several accounts, in the last 30 days. Each record in meta.fraud.device_profiles has device_id, profile_id, and document_count. To pull every document from that device, find device_id in meta.device_id and call:
Fraud history.
Raises the document score when the device has submitted fraudulent documents before. It looks back 30 days and needs at least 5 submissions. How aggressive it is can be tuned in Fraud Config.
Emulated device and blocked device.
Both are immediately red. Blocked device uses your own list of device_id values.
Document and Image signals
Handwritten characters
Handwriting on one or more extracted fields.
meta.handwritten_fields: all fields with handwriting. This is the combination of the two lists below. Check (meta.handwritten_fields) > 0 for any handwriting.meta.fully_handwritten_fields: fields that are fully handwritten. Lower risk.meta.handwritten_tampered_fields:printed values overwritten by hand. Higher risk.meta.handwriting_detected: true if handwriting is anywhere on the document, including outside extracted fields. On its own, it does not add a fraud type.
By default, Fraud Suite inspects total, subtotal, date, line_items.total, line_items.price, tax_breakdown.tax, and tax_breakdown.tax_base. You can change the list in Fraud Config.
Once handwriting tampering is enabled for the account, this type covers fully handwritten values only. Until then, handwriting over printed text also rolls into handwritten characters.
example #1
"fraud":
{
"attribution": "Handwritten characters",
"color": "red",
"decision": "Fraud",
"details": {
"handwritten_characters": {
"description": "Handwritten characters were detected on one or more
extracted fields."}
"types": [
"handwritten characters"
"fraud_review": {
"fully_handwritten_fields": [1 item
0: "line_items.3.total"
]
"handwriting_detected": true
"handwriting_tampering_present": false
"handwritten_fields": [1 item
0: "line_items.3.total"
]
"handwritten_tampered_fields": []
Handwriting tampering
Printed content overwritten by hand. This is its own fraud type once the signal is enabled.
meta.handwritten_tampered_fields:configured fields that were overwrittenmeta.handwriting_tampering_present: true if overwritten print is anywhere on the document, including outside those fields
Fraud Config has two modes:
fields(default): fires when a monitored field was overwrittenany: fires frommeta.handwriting_tampering_present, whether or not the mark is on a monitored field
"fraud":
{
"attribution": "Handwriting tampering",
"color": "red",
"decision": "Fraud",
"details": {
"handwriting_tampering": {
"description": "Printed content on the document was overwritten
by hand."
},
},
"digital_tampering_fields": [],
},
"types": [
"handwriting tampering"
],
"fully_handwritten_fields": []
"handwriting_detected": false
"handwriting_tampering_present": true
"handwritten_fields": []
"handwritten_tampered_fields": [1 item
0: "date"
]}
Digital tampering
Visible digital edits to a photo, such as changed numbers, text, or QR codes, or regions copied from elsewhere in the image. The list of edited fields is on meta.fraud.digital_tampering_fields, in dot notation such as line_items.1.total.
Note the path: it is under meta.fraud vs meta as handwritten.
By default, Fraud Suite inspects total, subtotal, date, line_items.total, line_items.price, tax_breakdown.tax, and tax_breakdown.tax_base. You can change the list in Fraud Config.
Not a document
The image is not a receipt or invoice.
"not_a_document": {"value": true, "score": 0.91}
value: the model verdict
score: the model's confidence.
This field meta.fraud.not_a_document is the raw result. The “not a document” type is the Fraud Suite decision after your threshold. Note: is_document on the document is a separate flag.
LCD photo
A photo of a screen, not of paper. Fraudsters often edit a receipt on a computer and then photograph the monitor. Read meta.fraud.pages[].is_lcd, which has value and score per page. The position in the array tells you which page.
Digital background (formerly screenshot)
The document started out digital rather than as printed paper: a screen capture, a PDF export, or another born-digital image. It is broader than a phone screenshot of an app.
"digital_background": {"value": true, "score": 0.98}
Value the model verdict
score: the model's confidence.
The digital background type only fires when the score also passes your Fraud Config threshold; if this signal is not configured, the response of this field will not affect color ad type but still be present in JSON.
During the rename, responses include both digital background and screenshot as types. Your logic should key off digital background and meta.fraud.digital_background.
Page detail. A separate page classifier says what kind of capture it was. It does not affect fraud color.
meta.pages[].screenshot.type: mobile_screenshot, other_screenshot, or null for a paper capturemeta.pages[].screenshot.score:that classifier's confidence
Generated document
The document was created with generative AI, or its provenance says it was produced by software rather than captured.
Visual check: meta.pages[].ai_generated with value and score.
Runs on images, not on text PDFs.
Metadata check: uses meta.source_documents[].exif when that metadata
is still on the file.
Note: Generated document is not reported for documents classified as "not a document", or for documents scanned or photographed with Veryfi Lens.
details.generated_document.reasons says why. For example:
ai_generated_model: The AI-generation model classified the document as generated.ai_generator_metadata:Document metadata identifies an AI generator.
A reason can include value, which is what the file declared about itself. Treat it as the file's claim, not Veryfi's.
Fraudulent PDF
Suspicious traits in a PDF, such as editing software in the metadata, overlapping text, or inconsistent fonts. The score is based on the strongest PDF signal.
Component scores are on meta.fraud.fraudulent_pdf. Every component key is always present:
null: that check could not answer0:the check ran and did not fire
details.fraudulent_pdf.reasons names what fired. For example:
suspicious_pdf_metadata:The PDF creator or producer metadata identifies suspicious document editing software. Includes field (creator or producer) and value.text_overlay:Overlapping, covered, or conflicting text was detected in the PDF.font_mismatch: Inconsistent fonts were detected across the PDF content.
Some reasons also include page.
Fraudulent image
Editing or re-encoding signals on a photo, not a PDF. Component scores are on meta.fraud.fraudulent_image.
This object is only returned when the signal is on for the account. If it is missing, the signal is off. Do not treat a missing object as a clean image.
details.fraudulent_image.reasons names the finding. For example:
editor_signature: The image names image editing software in its own metadata, so it was opened and re-saved by an editor. Includes field and value.double_jpeg: The image was JPEG-compressed more than once, so it is not the original photo or scan.
An editor signature is a reason to review, not an automatic reject. A user may have simply cropped a real receipt.
Duplicate
is_duplicate: probably the same receipt submitted again, based on a match of extracted field values
Similar Document
Similar documents (
meta.duplicates): text highly similar to earlier submissions. Each match has id, score, and url. Check (meta.duplicates) > 0 for any matches.
Before making a decision on whether to look at Duplicate or Similar Documents, read Duplicate vs similar documents.
Suspicious content
Mock, test, placeholder, or sample content on the document. This fraud type does not have a direct JSON signal yet, If enabled and configured, it returns Fraud Type suspicious_content
"suspicious_content": {
"description": "The document contains language or values commonly
associated with mock, test, placeholder, or sample content."
}
Data inconsistency (beta)
Receipt facts conflict with prior evidence, or a validation check failed. Off unless enabled in Fraud Config, and each check can be turned on separately. More details coming.
Other content checks
invalid qr data: the QR payload does not match the extracted fields
vendor layout mismatch: the layout does not match the expected template for this vendor
total in words mismatch: the written amount does not match the numeric total
other: a fraud signal was detected that does not have its own type
These return a description in meta.fraud.details and have no raw field or score.
What details look like
meta.fraud.details has one key per fired type. Every fired type has a description. Types with reasons also return a reasons list. Each reason has a stable name and a description, and sometimes the field, value, or page that triggered it.
"details": {
"digital_background": {
"description": "Digital-background artifacts were detected in
the document image."
},
"fraudulent_pdf": {
"description": "The PDF contains indicators of editing or manipulation.",
"reasons": [
{
"name": "suspicious_pdf_metadata",
"description": "The PDF creator or producer metadata
identifies suspicious document editing software.",
"field": "creator",
"value": "Example Editor"
}
]
}
}
Types that return reasons: generated document, fraudulent pdf, fraudulent image, data inconsistency, and the legacy screenshot key. All other types return a description only.
New reason names can be added over time. Build your logic so an unknown reason name is shown to the reviewer rather than dropped.
Warnings
meta.warnings flag calculation and integrity issues, such as line items that do not add up to the subtotal. They can have legitimate explanations, so treat them as warnings only. They are not fraud types and they do not affect color, score, or types. If you build your own logic, you can use them as extra inputs.
Each warning has a type and a message.
On by default:
line_item_amount_mismatch
line_item_repeats
item_counts_mismatch
totals_mismatch
tax_rate_mismatch
barcode_decoding_issue
barcode_code_missing_in_ocr
logo_vendor_mismatch
malware
Off by default:
weekend_transaction
time_and_currency_mismatch
exif_creation_modified_date_mismatch
missing_date
suspicious_document_aspect_ratio
How to get access
Fraud Detection Suite is a premium feature and needs to be activated on your account. Contact your Veryfi account manager or Sales to enable it. If you need a signal that is not listed here, your account manager can help with that too.
How to configure fraud
Configure fraud signals yourself under Settings > Fraud Config. Turn each signal on or off, set its sensitivity, and choose the color it can force.
Step-by-step guide: Configuring fraud signals in Veryfi.
Fraud signals analytics
See how often each signal fires across your documents under Analytics > Fraud Stats. Use it to check your thresholds before and after you change them.
Review flagged documents in the Inbox
Use the fraud filters in the Inbox to find and review flagged documents.
You can find it at: Inbox>Fraud Filter
How can I help improve Veryfi's fraud detection accuracy?
When your team reviews a document, send the outcome back. Your decisions help Veryfi tune detection for your traffic.
Update the document (PUT) with meta.fraud_review:
decision: fraud, not fraud, or unknown
types: required when decision is fraud. Must be empty otherwise.
{
"meta": {
"fraud_review": {
"decision": "fraud",
"types": ["digital tampering"]
}
}
}
Types you can send: handwritten characters, handwriting tampering, digital tampering, generated document, LCD photo, digital background, not a document, duplicate, high velocity, critical velocity, multiple profiles or devices, fraudulent pdf, fraudulent image, invalid qr data, suspicious content, data inconsistency.
The legacy names ai generated and screenshot are still accepted. Use the current names for new integrations.
API reference: Update a document.
Related articles
API docs: Fraud overview, Signals, Triggers, Fraud types
Need help? Contact support@veryfi.com






