Video Games Analysis for Personal Data Prediction

The aim of this thesis is to try to use in-game generated data in order to predict players' private information, such as gender, age, economic state, and personality traits. In contemplation of such purpose we propose a machine learning-based attribute inference attack, assessing its feasibility for the specific case study of Dota 2. Therefore, we dealt with data collection of consensus players' gaming and private information, creating a dataset to employ for our attack. Starting from that, two separated analysis were pursued to find which parameters are correlated with sensitive data: one taking into consideration single matches of the game and the other accounting for players' overall statistics. At both analysis levels, significant correlations between the private features of interest and the gathered attributes are reported and exhaustively commented before actually performing the predictions. Despite the employment of several models and the application of machine learning good practices, we were not able to obtain useful classifier for all the studied target features. However, we indicated some expedients and interesting ideas to continue the research in this promising ambit. We then proceeded with our dissertation by discussing on the applicability of the suggested attack to other games and reasoning on the limitations of our study. Concluding that, despite the obtained results, we are confident in declaring that this attack has a really powerful potential, since correlations between data generated by players' when gaming and their private information are actually there, just waiting to be exploited.