Anti-cheating

[scchess]

I'm looking for people who can help with anti-cheat detection or fraud detection in chess. The aim is to find out who has cheated in chess with an engine, while constructing a robust statistical framework to control the false positive rate like chess.com/lichess has achieved. It's going to be a paid job. Anyone interested or knowing someone who has expert knowledge on it, please DM me. Thanks

[carldaman]

Hi Ted,
Please check your PM inbox

[towforce]

This is "friendly discussion" rather than "expert help", but isn't detection of cheating mainly down to absence of shallow tactical errors?

[Cornfed]

This is "friendly discussion" rather than "expert help", but isn't detection of cheating mainly down to absence of shallow tactical errors?

From what I have read it is more than that. And think about it - slower time controls produce few (if any) 'shallow tactical errors' in a game or over the span of several and would provide the cheater with a pretty easy means of avoiding detection.

[jdart]

Really, the performance history tells you everything. On the chess servers, anyone who's winning regularly against computer opposition is cheating. Arasan some time ago played a series of games against a 2500+ GM and the GM lost all of them. Of course most players are lower-rated and will have even less chance. So just look at the rating and the win/loss record.

[scchess]

To me, anti-cheating is more than just checking engine similarity to played moves. It's what we call "fraud detection", very common in insurance and many other fields. Fraud detection takes many input features, engine similarity is simply just one of those features. Other features can be:

* Performance history
* Ability to not to blunder
* Correlations to the top N-th PV moves
* Ability to play well in high complex positions and its correlation with time-management
* etc etc.

We'd then form a statistical model such as frequentist and Bayesian from the features. If we use frequentist, the model would then tell us the probability of someone is cheating, assuming a null hypothesis no cheating occurred. We'd then control the false positive rates with something q-values.

Anti-cheating is much more than just computer engines PV scores.

[scchess]

Really, the performance history tells you everything. On the chess servers, anyone who's winning regularly against computer opposition is cheating. Arasan some time ago played a series of games against a 2500+ GM and the GM lost all of them. Of course most players are lower-rated and will have even less chance. So just look at the rating and the win/loss record.

True, but what about players who cheat from FIDE 1000 to FIDE 1500?

[scchess]

This is "friendly discussion" rather than "expert help", but isn't detection of cheating mainly down to absence of shallow tactical errors?

Certainly not, anti-cheating requires a proper statistical model.

[scchess]

What I imagine, anti-cheating works like the dashboard we can all get from your own "Chess Insights" on lichess. Try it yourself. You will see lot's of statistics from your own games such as centipawn loss bucket. Lichess would of course know everybody's centipawn loss bucket. They could then aggregate all the statistics to form an average for each rating group.

From the number of games you play, it should be possible to form a statistical confidence how likely you are really in the rating group.

Now, if for any reason suddenly, your "centipawn loss bucket" outperforms most of the people in your rating group especially during a lichess tournamenet, they would know you cheat even you may not actually win the games (i.e. loss on time). Your WDL remains but they know you cheat.

[cyrill57]

I think the interesting idea is to train a neural network to detect engine/human games. After you train it on a large enough dataset, it will be relatively easy to check statistics for false positives for such a network.