Counterfeit detection vs authentication: why computer vision can assess but not confirm

Image-based authentication has become a standard answer to counterfeits. Photograph the stitching, the heat stamp and the grain of the leather, and a model trained on thousands of genuine and fake examples returns a verdict, often with a percentage beside it. It is fast, it scales, and for many items already on the market it is the best check available.

It is also answering a narrower question than the one a brand, a reseller or a buyer is asking. A check like this assesses whether an item looks like genuine examples of its model. It cannot confirm which item it is. That difference decides what each kind of check can be trusted with.

A classifier compares the photographs in front of it with what it learned from other items: genuine pieces of the same model, or stock-keeping unit (SKU), and the fakes it was shown. Its answer is a judgment about membership in a class. This bag is consistent with authentic examples of this style. It is not a statement about this bag, because the model has never seen it before and holds no record of it.

Two things follow. First, a genuine item with the wrong story passes: a bag reported stolen, a watch sold through a channel the brand never approved, a piece rebuilt with parts from other genuine pieces. Each looks like its model because it is its model. Second, the percentage describes the classifier, not the item. Research published in 2017 found that modern neural networks are poorly calibrated: the confidence they report and how often they are actually right can drift apart.¹

Even a perfectly calibrated check depends on something it cannot see: how many fakes are in the pile. Take a hypothetical check that is right 99 times in 100, on fakes and genuine items alike, used where 2 in 100 of the items sent to it are fake. Of 10,000 items, 200 are fake, and 198 of those are flagged. Of the 9,800 genuine items, 98 are flagged by mistake. Of the 296 items flagged, about 1 in 3 is genuine.

The same check, with the same accuracy, means something different in a channel full of fakes than in one with almost none. A likelihood is a statement about a population. A buyer holding one bag wants a statement about that bag.

Counterfeiters do not stand still. In 2025 the U.S. Chamber of Commerce described today's superfakes as nearly indistinguishable from the real thing, and some as convincing enough to fool experts. A superfake is made to pass exactly the checks that buyers, resellers and authenticators use.²

An image model is measured on the fakes that existed when it was tested. Machine-learning research shows how fragile that measurement can be. When researchers rebuilt the ImageNet test set by following the original process, the same models lost 11% to 14% of their accuracy. Fakes made to pass are a harder shift than a new test set, because someone is working against the check. Images altered to make a classifier misread them have kept working even when seen through a phone camera.^3,4

This is the sense in which image-based authentication assesses rather than confirms. Each season of better fakes lowers what its verdict is worth, and the check cannot tell which season the item in front of it comes from.

Confirming identity is a different test. Instead of asking whether an item resembles its model, it asks whether the item matches a record made for that one item when it was made. That comparison does not weaken as fakes improve. A copy can resemble the original perfectly and still fail to match a record it was never part of.

Imaging can do this too, when the record comes first. In 2002 physicists showed that the microscopic structure of a disordered material can be reduced to a unique identifier that is hard to duplicate. In 2005 a team showed that paper, plastic cards and product packaging carry a unique fingerprint in the microscopic imperfections of their surface, readable with a low-cost laser scanner. Both work the same way: record the item first, then check it against that record. The fingerprint confirms identity only because it was captured at origin.^5,6

Regulation is moving toward the item too. Under the EU's Ecodesign for Sustainable Products Regulation, a product's passport is reached through a data carrier on the product, and each delegated act decides whether the passport is kept per model, per batch or per item. Only the item level can say which product this is.⁷

None of this makes image checks pointless. Most products in circulation were made before anyone recorded them: a vintage bag, a watch from the 1990s, stock already in the market. For those, assessment by trained eyes and trained models is the best evidence there is, and a careful check is far better than none.

The mistake is treating an assessment as a confirmation, or resting a brand's protection on a test its counterfeiters can study. For new production the order can be reversed. Record each item where it is made, and every later check becomes a comparison with that record rather than a judgment of resemblance. Computer vision is a tool to assess. Confirming an item takes a record that was there first.

A tag does the same job as a recorded fingerprint, by other means. A Vera tag is fitted where the product is made and bound to a certificate that belongs to that one item. Each time it is tapped, the tag computes a fresh cryptographic code that is checked against the record. A copied link opens the certificate without the verified state, and a replayed code is recorded as invalid. From the first record on, the certificate carries who owns the item now and every transfer, resale and scan, from the factory to the latest owner.

That is the line between a likely match and a confirmed one. An image check can tell a buyer that a bag looks right. A record made where it was made can tell them which bag it is, and whose.