How Can Marketplaces Detect Impersonation and Synthetic Media Fraud?

How marketplaces detect impersonation and synthetic media fraud

Marketplaces catch synthetic media fraud by screening media at each checkpoint where a fake can be cashed out, rather than once at signup. A seller who passes onboarding with a real identity can still be taken over months later, and a genuine buyer can submit an AI-edited photo to win a refund. Detection has to follow the money.

At each checkpoint the pattern is the same. The submitted selfie, video or photo goes through manipulation and generation analysis, the result joins device, network, account age and transaction signals in the risk engine, and the policy decides whether to approve, step up or send the case to a reviewer. The table lists the checkpoints and what each one protects.

Which synthetic media fraud patterns hit marketplaces?

Synthetic sellers are the most damaging pattern. A fraudster creates a seller account with a generated ID image and a matching synthetic or face-swapped selfie, lists high-demand goods, collects payments and disappears before buyers complain. 404 Media reported in February 2024 that an underground site called OnlyFake sold AI-generated ID images for $15 each and that a reporter used one to pass identity verification at a cryptocurrency exchange.

Impersonation of real people follows. Attackers clone the face or voice of an established seller, a brand representative or a creator to pass re-verification, change payout details or run fake endorsements. The FBI's public service announcement of 3 December 2024 describes criminals generating realistic photos for fake social media profiles and short videos to convince victims an online contact is real.

Refund fraud is the quietest pattern. A buyer edits a delivery photo to show a crushed box or an empty package, or generates an image of a damaged product, and files a claim. Each claim is small, so these files rarely get careful review. The workflow guide on impersonation controls for marketplace refunds covers how to route them.

What the INFORM Consumers Act and DSA Article 30 require

Both the US and the EU now require marketplaces to verify the sellers they host, which makes a convincing fake ID or selfie a compliance problem as well as a fraud problem.

Under the INFORM Consumers Act, in force since 27 June 2023, a high-volume third-party seller is one with 200 or more sales of new or unused products and $5,000 or more in gross revenue in any continuous 12-month period within the past 24 months. The marketplace must collect bank account, tax ID and contact details within 10 days of a seller qualifying, verify them within 10 days of receipt, and require sellers to certify their information at least once a year. For entities, that includes a government-issued ID for the person acting for the seller or a record showing the business name and address.

Article 30 of the EU Digital Services Act requires platforms that let consumers buy from traders to obtain the trader's name, address, contact details, a copy of an identification document, payment account details and trade register number where one exists. The platform must make best efforts to check the information is reliable and complete, and suspend traders who fail to correct inaccurate data. The seller identity controls guide for marketplace onboarding maps both rules to an onboarding flow.

Building the checks: a practical order of work

Most marketplaces cannot screen everything on day one. Start where losses concentrate and where a single fake opens the door to many transactions:

• Add face image and video analysis to seller onboarding and to every payout-detail or contact change, since those are the steps that move money.
• Bind re-verification to the face captured at onboarding, so a deepfake of a different person fails even if it looks real.
• Screen evidence photos and videos on refund claims above a value threshold, and route flagged claims to a trust and safety reviewer.
• Feed detection scores into the existing risk engine alongside device, account age and chargeback history, as FinCEN suggests for new accounts with high chargeback volumes.
• Review the flagged queue weekly and adjust thresholds against confirmed fraud and false positives.

Where detection helps a marketplace and where it does not

Detection is strongest on media that was generated or edited: synthetic faces, face swaps, inpainted damage on a product photo or a pasted shipping label. It gives reviewers a reason to look closer and evidence to cite when they deny a claim.

It cannot confirm that a genuine, unedited photo shows the item actually delivered, that a real person is not acting as a paid front for a fraud ring, or that a seller's tax details are true. Those questions need order data, logistics scans, document and registry checks and account history. A detection score is one input to the decision, and the marketplace's policy makes the call.

How Neural Defend fits marketplace trust and safety

For seller onboarding and re-verification, NeuroVerify checks a face image, video or audio clip in one API call across six signals, including Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis and AI Pixel-Level Manipulation. It returns one unified risk score with signal-level evidence in under 200 ms p95 for eligible image requests under defined test conditions, fast enough for a signup step. SDKs cover Android, iOS, Flutter, Python, TypeScript, Go and Java.

For refund disputes and investigations into seller rings, NeuroDetect lets a trust and safety analyst upload evidence photos, videos or documents into a case, inspect heatmaps down to specific regions and frames, and export a report that records uncertainty and model version. That report is useful when a seller or buyer appeals.

Both build on ARCAS 1B, Neural Defend's own detection model. On the public VendorBench-100 benchmark of 100 adversarial still images, including AI photo edits, Neural Defend ranked first of 36 systems with an MCC of 0.876 against 0.646 for the next vendor. One VendorBench author also co-authored the ARCAS paper; the corpus and results are public so anyone can rerun them.