AI could identify anonymous chess players and pose a privacy risk

With the advent of AI that exceeds human decision-making ability in complex areas, AI has come to surpass humans in shogi, go, chess, strategy games, etc. Meanwhile, AI that can identify the player with high accuracy from the chess piece movement has been announced, and it is feared that privacy in online chess may be threatened.

Detecting Individual Decision-Making Style: Exploring Behavioral Stylometry in Chess
(PDF file)

AI unmasks anonymous chess players, posing privacy risks | Science | AAAS

In 2021, a research team at the University of Toronto announced a technology to identify individuals from chess playstyles using machine learning models. The research team prepared more than 50 million game records of players who played at least 1000 times on the online chess site Lichess, sampled up to 32 moves from them, coded them, and let AI learn them. Then, 100 game records from each of about 3,000 players and 100 new game records played by one player selected from about 3,000 players were analyzed by AI with the first 15 moves hidden. I was allowed to.

As a result, AI seems to have succeeded in identifying the selected players with 86% accuracy. In addition, when similar inferences were made without using AI, the accuracy was only 28%, so you can see that AI's analysis accuracy is high.

Ashton Anderson, the principal investigator of the research project, says that most chess AIs play in a 'foreign style' that is not very useful for people looking to learn or improve their skills. Mr. Anderson said that learning and improving chess skills requires advice tailored to each player's play style, and that AI needs to capture the player's unique form.

However, Reid McRoy-Young, who belongs to the same research team. We are aware of the risk of privacy violations in this AI, and point out that it may be used to uncover the identity of anonymous online chess players. McRoy-Young says that if you adjust the AI, you can do the same with poker, and theoretically, if you can prepare a data set, you can identify people from their driving habits and when and where they use their mobile phones. .

``Privacy threats are rising rapidly,'' said Alexandra Wood, an attorney at the Berkman Klein Center at Harvard University. In that sense, the results of this study are useful.

The organizers of the Neural Information Processing Systems Conference (NeurIPS), where the research was published, believe that the study of identifying players with chess AI is technically brilliant but ethically problematic, and details privacy risks. He said that he accepted the research presentation on the condition that he would do it.

In addition, Mr. Anderson has announced that he will not disclose the source code of AI.

in Note, Posted by log1i_yk