[salgernon]

I always thought that a near learning project would be training an ML on “real” cards and then detecting fakes. I don’t play the games but I was always thrown by how much effort went into counterfeits, but I guess there’s enough profit for someone. There’s usually something wrong with the registration or colors.

[FanaHOVA]

What is missing in the context here is that the cards mentioned in this article are not actually real. They never existed, and therefore they are not "counterfeits" of a real one, they are just made up. Someone just claimed to know someone that had playtest cards from back in the day. They are not a commercial product.

See here for a bit more background: https://www.cgccards.com/news/article/13347/

[strstr]

If you are willing to pull out a loupe you don’t really need ML. You can just look at the rosette patterns.

For Mtg cards, the green dot test is very easy to learn, and I’m not familiar with any fakes that pass it.

(Edit: arguably you have to worry about rebacking with the green dot test, but rebacking is typically pretty fishy looking.)

[nilamo]

Pulling out a loupe and manually inspecting a card is a slow process if you have a few thousand cards (avg player).

[FanaHOVA]

Avg player doesn't buy a few thousand cards at a time. If you buy a high value card from a random seller you should always check it unless you trust them from references.

[strstr]

People only pull out slower tools for valuable, forgery worthy cards.

If someone is buying 1000 $1000 dollar cards, it’s still worth it lol.

Even cheap forgeries cost money to produce, so I wouldn’t expect a lot of low value cards to be forged. If you sort out the valuable cards and do random sampling, you can probably catch the most problematic cases.

[krisoft]

> There’s usually something wrong with the registration or colors.

That can be selection bias too.

Maybe the counterfeits where there is nothing wrong with the registration of colours are just not recognised as counterfeits.

Similarly how seemingly every hacker you can hear about in the news are bad at opsec. Because you wouldn't hear about them if they weren't.

[HanClinto]

I built one of these several years ago for MtG cards. Trained a neural network with a binary classifier on a cheap $20 USB microscope looking at examples of the backs of real cards vs. fake cards.

https://youtu.be/6_kKR7YgPF4

Sadly never got around to shipping it, because it worked really well. Ported it to the web, but never figured out the billing issue, and so it died during the delivery phase. From time-to-time, I still wonder if I should resurrect this project, because I think it could help a lot of people.