Image Forensics: Detect Edited & Fake Images

Veryfi Image Forensics reads every layer inside a receipt photo, a proof of purchase, or a claim photo: pixels, error levels, camera noise, EXIF metadata, and file structure. It checks them against 110 fraud signals in under 3 seconds and returns a risk score, the evidence, and a plain-English explanation before the image enters your workflow.

Get Started for Free Free Demo
Image Forensics: Detect Edited & Fake Images

Part of the Fraud Detection Suite

These checks run with the rest of Veryfi Fraud Detection and come back in the same response, under meta.fraud: one color, a score, and the reasons.

See every fraud check
  • AI-Generated Documents

    Spots receipts, invoices, and checks made with ChatGPT, Midjourney, Stable Diffusion, DALL-E, or GANs.

    You are here
  • PDF Forensics

    Reads inside the file: the app that made it, its fonts, and text laid over the original. Flags edited and AI-generated PDFs.

    You are here
  • Image Forensics

    Reads the layers inside a photo: error levels, camera noise, copied pixels, and EXIF data. Flags edited, AI-generated, and recaptured photos.

    You are here

How Veryfi reads the layers inside a PNG image and uses multi-modal AI to detect fraud

A PNG photo of a receipt looks like one picture, but Veryfi reads it as eleven layers: the rendered image, the characters, error levels, sensor noise, copied pixels, JPEG compression history, the EXIF thumbnail, EXIF data, text chunks, and the file structure, and then a multi-modal AI model reads the pixels, text, and metadata together. In this example the total was changed from $48.60 to $148.60. The photo looks normal, but nine layers show the edit: a digit that sits off the printed baseline, a bright spot in error level analysis, a patch with no camera noise, pixels copied from another line, a break in the JPEG block grid, an EXIF thumbnail that still shows the original total, EXIF dates that disagree, text chunks written by a photo editor, and an AI model that scores the image 3% likely to be AI-generated but 97% likely to be edited at the total. The file itself is a valid PNG. Veryfi Fraud Detection reads every layer, not just the pixels, and uses AI to tell a generated image from an edited one.

AI image forensics

The photo looks real. The layers say otherwise.

Veryfi Inspect X-Ray checks every receipt photo against 110 fraud signals

A photo of a receipt goes into Veryfi Inspect X-Ray and is read against 110 fraud signals in three families. File and origin (34 signals): 14 file signals, 8 metadata signals, and 12 timeline and duplicate signals. Pixels and AI (25 signals): 16 pixel signals and 9 AI-generated and screen recapture signals. Text and PDF (51 signals): 9 OCR signals, 26 PDF structure signals, and 16 PDF content signals. On top of these, 14 document integrity checks confirm the file is complete and consistent. Out come three results: the extracted data (merchant, date, items, tax, and total), the fraud signals that fired (a photo of a screen, pixel edits around the total, a camera date after the purchase date, and a possible duplicate of an earlier submission), and an insight (a risk score of 87 out of 100, so the receipt is held for review).

Vee, the Veryfi Inspect advisor, explains every fraud signal on a receipt photo and what to do next

Vee Advisor

Every signal, explained by Vee.

  1. Hi, I’m Vee. I read this receipt photo against all 110 Inspect X-Ray signals. Let me walk you through what fired, why it matters, and what to do next.
  2. Scan Signals: 94% risk. 9 of 110 signals fired across four families, and they all point at one field: the total. That’s high risk.
  3. Verdict: reject the expense as submitted. I’m confident: the pixels, the math, and the metadata agree. The rest of the receipt is real, so if you pay anything, pay the original $48.60.
  4. 1. Total edited. Someone pasted a “1” in front of $48.60. One digit adds $100 to the claim.
  5. Zoom in on the total. The “1” glows at error level 41 while the rest of the photo sits at 3 to 6. It has no camera grain, and it matches the “1” in item 2 at 97.9%.
  6. 2. Total doesn’t add up. Subtotal, tax, and tip come to $48.60. The printed total is $148.60, off by exactly $100.
  7. Here are the numbers. 39.35 + 3.25 + 6.00 = 48.60. Every line reads cleanly, so this is not an OCR slip.
  8. 3. Edited after capture. The camera saved this photo on Nov 2. An online photo editor saved it again on Nov 8.
  9. Here’s the date. The receipt and the camera both say 11/02 at 12:41. The file was saved again on 11/08 at 22:15, and the camera’s own preview still shows $48.60.
  10. Saved again 6 days later. The photo was taken at the restaurant on Nov 2, edited on Nov 8, and submitted the next day.
  11. Tax checks out, the total doesn’t. 8.25% of 39.35 is 3.25, so the original receipt was consistent. Only the total moved.
  12. Got questions? Chat with me. Ask anything about this receipt. I answer from its X-ray report, like whether the camera app could have done this.
  13. That’s every signal. 3 findings: 2 high, 1 medium. Reject it as submitted and ask Dana for the original photo or the card statement line.
JSON Example
“meta”: {
“duplicates”: [
{
“id”: 191347946,
“score”: 1,
“url”: “https://scdn.veryfi.com/receipts/…”
}
],
“fraud”: {
“attribution”: “High velocity, Handwritten characters, LCD photo, Digital background, Duplicate receipt”,
“color”: “red”,
“decision”: “Fraud”,
“digital_tampering_fields”: [
“total”,
“line_items.10.total”,
“line_items.14.total”
],
...
JSON Example
...
“images”: [
{
“is_lcd”: true,
“score”: 0.96
}
],
“score”: 1,
“types”: [“handwritten characters”, “LCD photo”, “screenshot”, “duplicate”, “generated document”]
},
“handwritten_fields”: [
“line_items.0.total”,
“line_items.1.total”
],
“warnings”: [
{
“message”: “Line item #14 has a problem. Quantity x Price != Total”,
“type”: “line_item_amount_missmatch”
}
]
}

Developer Friendly API (Headless Option)

Image fraud signals are baked into the same JSON response as your extracted data API. 1 API. No second API call. No extra integration. AI that reads every photo and knows when it’s been edited.

Most teams are up and running in an afternoon.

color is your routing logic.
attribution is your audit trail.
mages.is_lcd tells you the photo was taken of a screen.

Thresholds ship with sensible defaults. Tighten or loosen per signal whenever you’re ready.

Use Cases

Anywhere a photo is proof of money spent, damage done, or who someone is. See how Veryfi reads the layers inside a photo.

Frequently Asked Questions for Image Forensics

  • What is image forensics?

    It means checking the inside of a photo (pixels, compression history, camera noise, EXIF metadata, file structure) to find out whether it was changed after the camera took it, or never came from a camera at all. Looking at the photo can’t show this.

  • How is this different from OCR or a visual check?

    OCR reads what the receipt says. Forensics checks whether what it says is original. An edited photo can look perfect and still carry a patch with no camera grain, a digit copied from another line, or a photo editor’s name in its metadata.

  • Which images does it work on?

    Photos and scans of receipts, invoices, IDs, bills, and other documents in JPEG, PNG, and other common image formats, including screenshots.

  • Does it catch a photo of a screen?

    Yes. Screen recapture has its own signals. A photo of a monitor or phone comes back flagged as an LCD photo, with `is_lcd` and a score in the JSON.

  • Can it tell an AI-generated image from an edited one?

    Yes. The multi-modal model scores both separately, for example “3% AI-generated, 97% edited at the total.”

  • Which tools does it catch?

    Photoshop and other photo editors, online photo editors, phone markup apps, receipt generator websites, and images made by ChatGPT or other AI image tools.

  • What if the EXIF data was removed?

    Removing metadata is itself a signal, and the pixels (error levels, camera noise, copied regions, compression history) still carry traces of the edit.

  • Will it flag real photos just because they were compressed or sent through a messaging app?

    Not on its own. One signal changes the score, but red needs several signals pointing to the same field. Yellow sends the image to a person for review.

  • Does it catch the same receipt submitted twice?

    Yes. Each image is compared with earlier submissions, and matches come back in the “duplicates” list with a score and a link to the earlier one.

  • Do I have to set up rules or train a model?

    No. It works on the first day, and you choose what green, yellow, and red do in your workflow.

Fraud is live 24×7

Intercept Document Fraud

Intercept forged IDs, padded expenses, and fake invoices flooding KYC/KYB, Travel & Expense, Bill Pay, Audit and other areas where fraud is a risk. Stand up Veryfi in minutes and shut them out before they clear.

No credit card required. Live in minutes.