Seller and refund fraud detection for marketplaces
Marketplaces take a lot on trust from pictures: the shop and the documents at seller onboarding, the product photos in a listing, and the photo of a damaged item in a refund claim. TrueSecurix checks each image and tells your trust and safety team whether to pass, review or flag it, with the evidence.
In our held-out test of 31,966 images, 0 of 2,526 genuine business photos (shops, markets, receipts, documents) were flagged and 87.8% of manipulated images were caught. Read how we measured
What trust and safety teams at marketplaces are seeing now.
What seller and refund fraud looks like today
Sellers with no shop
A generated storefront or warehouse, or a real one with the signboard renamed to match the application.
Edited onboarding documents
A registration certificate, a cancelled cheque or an ID photo with a name or a number changed by hand.
Listings with no product
Product photos made by an image model, for goods the seller does not hold.
Damage added for a refund
A crack added to a phone screen, a stain to a garment or a dent to a parcel by an AI editor, to claim a refund or a replacement.
Image and identity checks run on the same account and the same credits: one credit per check.
Which checks apply, and what you get back
Each image goes to POST /v1/image/authenticity; selfies, IDs and statements go to POST /v1/verify. Every response carries an outcome and the evidence behind it, and every check is written to your history.
| What you receive | Check | What the response gives |
|---|---|---|
| Shop and warehouse photos at onboarding | Image authenticity | Real, AI generated, AI edited or digitally edited, with the indicators and the located edited areas. |
| Business documents, as photos or scans | Image authenticity | The same four answers, with the edited areas located so a reviewer knows where to look. |
| Seller PAN, Aadhaar and selfie | Identity checks | Checksum validation of the numbers, deepfake selfie detection and face match to the ID photo. |
| Listing photos | Image authenticity | Whether a product photo was generated by an image model, with the evidence. |
| Refund and damage-claim photos | Image authenticity | Whether the damage was added by an AI editor or by hand, with the edited areas located. |
Nobody is rejected on a guess. The model tells you which files a person should look at.
Three outcomes your trust and safety team can act on
The decision field turns the answer into an action. Route on it directly in your trust and safety case tool.
Accept it
No sign of generation or editing. The seller, listing or refund continues through your normal checks.
A person decides
Some evidence points to a manipulation, not enough to accuse a genuine seller or buyer. An agent takes a short look, guided by the class it leans toward.
Treat as manipulated
Strong evidence of generation or editing, with the indicators and the edited areas marked. Hold the onboarding or the refund and ask for more evidence.
How a team rolls it out
Most of the work is wiring three outcomes into the trust and safety case tool you already run. You can build all of it before you spend a credit.
One request, one image, one answer
curl https://truesecurix.com/v1/image/authenticity \ -H "X-API-Key: $TRUESECURIX_API_KEY" \ -F "image=@damaged_item.jpg" \ -F "reference=refund-77310"
Field by field in the API reference.
- Build against a test key. A test key returns a simulated answer at no charge, including the exact review shape when you send
reference=review, so your trust and safety case tool handles all three outcomes before any live image is sent. - Run a proof of concept on your own labelled images. A folder of real and manipulated seller, listing and refund photos from your own files, with a CSV of labels, goes through the batch script. You get accuracy per class, the confusion matrix, genuine images wrongly flagged and the share sent to review, on your data.
- Send results into your trust and safety case tool. Signed webhooks post
check.flagged,check.reviewandcheck.completedto your HTTPS endpoint, signed with HMAC-SHA256 and retried for about a day. - Route reviews to a person. A review carries the class the evidence leans toward and the indicators behind it, so a reviewer knows what to look for. Every check stays in your history with its model version, and exports to CSV.
Measured on images it never saw
The image model was measured on business photos of shops, markets, receipts and documents, plus general photos, in a held-out test of 31,966 images never used in training, threshold setting or model choice.
Your own images are the measure that counts for your use. A proof of concept on your own labelled images gives you the same figures on your data, with the per-image CSV behind every number.
Questions risk, operations and engineering teams ask first.
Questions, answered
Which seller and buyer images can it check?
Any JPEG, PNG or WebP image up to 15 MB: shop and warehouse photos and business documents at onboarding, listing photos, and photos sent with refund or damage claims. Seller IDs and selfies go through identity checks on the same credits.
Will it flag listing photos that a seller has retouched?
A photo that was retouched in an image editor can come back as digitally edited, because it was edited. Decide in your own policy whether that matters for listings. For refund and damage claims, where the photo is the evidence, it usually does.
Does it block sellers or refunds automatically?
No. Each image comes back as pass, review or flag with the evidence. Your policy decides what happens next, and a review always goes to a person.
How accurate is it on our images?
The model was measured on business photos of shops, markets, receipts and documents, plus general photos, in a held-out test of 31,966 images. It has not been measured on marketplace images specifically. A proof of concept on your own labelled images is how you measure it for your use.
Check seller and refund photos before you act
Try a photo in the live demo now, or create an account for 100 free checks. No card is needed, and test keys are free while you build.
Running lakhs of checks a month? See enterprise terms and a proof of concept.