Neural net for cheat detection

[Fulvio]

If one have a big database (let's say lichess) and a list of the players banned for cheating, it is possibile to train a network that will predict suspicious games?

[j.t.]

If it worked, then cheaters could use such a network to avoid suspicious moves.

[AdminX]

If one have a big database (let's say lichess) and a list of the players banned for cheating, it is possibile to train a network that will predict suspicious games?

I am sure someone out there must be working on this, even if not for chess. However, I think it might always be a case of you build a better mouse trap, and they will find / train a smarter mouse.

[Damir]

As Elon Musk said: the so called cheaters, ''they have anal beads up their ass''.

[chrisw]

If one have a big database (let's say lichess) and a list of the players banned for cheating, it is possibile to train a network that will predict suspicious games?

This hypothetical LiChess data will be full of false negatives and false positives, ie full of labelling errors. Neural nets are at best approximators (ie full of errors), so you would just end up with errors on top of errors - not very useful.

[scchess]

It's not that simple... I worked for chess chess detection as part of my works...

[scchess]

If one have a big database (let's say lichess) and a list of the players banned for cheating, it is possibile to train a network that will predict suspicious games?

This hypothetical LiChess data will be full of false negatives and false positives, ie full of labelling errors. Neural nets are at best approximators (ie full of errors), so you would just end up with errors on top of errors - not very useful.

Correct. In cheat detection, training a model is easy. It's all about data processing. Generally, neural network is not as useful as other simpler models for cheat detection because of it's black-box model, high likelihood of overfitting, and more works be done for hypothesis testing.

[scchess]

In contrast to everybody thinks, cheat detection at chess is not just a data science problem. Data Science is simply part of the whole process. It's a complicated area that involves strong engineering support, statistical modeling, ML ensemble models, data processing, management of moderators etc. Neural network will never able to resolve cheat detection by itself. To make cheat detection working, technical skills are not enough. Soft skills especially good management skill of must be part of the equation.

[dkappe]

If one have a big database (let's say lichess) and a list of the players banned for cheating, it is possibile to train a network that will predict suspicious games?

If you look at the statistical model developed by Regan, it essentially predicts how often a strong human player agrees with an engine. If a player deviates from that, you have statistical evidence that he used an engine.

You don’t need a neural net for that sort of thing.

[scchess]

If one have a big database (let's say lichess) and a list of the players banned for cheating, it is possibile to train a network that will predict suspicious games?

If you look at the statistical model developed by Regan, it essentially predicts how often a strong human player agrees with an engine. If a player deviates from that, you have statistical evidence that he used an engine.

You don’t need a neural net for that sort of thing.

The thing is ... Regan's model is practically useless for online cheat detection... I rejected it for good reasons and opted for an alternative. His model is more like a reference guide, and is the model used by FIDE. It has several clear disadvantages:

• The model is unable to account for non-chess engine metrics. Clearly doesn't work for online cheat cheating.

• The model is unable to account for smart cheaters

• The model is unable to account for slow and gradual rise in ELO rating

Regan's model was made for OTB games where the metrics were small and simple. I