Detecting Tool-Assisted Cheating in Competitive Video Gaming Using Machine Learning
Open Access DepositedCheating in competitive video gaming undermines fair competition, jeopardizes the integrity of eSports and has a direct, material impact on game developers’ revenue and cost. Anti-cheat software is, typically, closed-source and must be run locally on the player’s machine with kernel-level access. New cheating methods are being developed that provide players with tools that run externally from the game and, even, the computer running the game, making traditional anti-cheat tactics less effective. This praxis demonstrates the creation and application of a machine learning model to detect tool-assisted cheating without requiring direct access to the players’ machine through the analysis of player input and game state features in a racing game without requiring extensive game-specific domain knowledge. The study develops and documents a process for building an extensive dataset comprised of real-world data obtained from publicly-available sources. The author obtains hundreds of millions of frames of real player input, transforms and enriches the input with additional game state data from the game. Preprocessing steps include data cleaning, feature selection, and augmentation strategies to mitigate class imbalance. Among the evaluated models, CatBoost demonstrated the best performance in distinguishing between unassisted and tool-assisted sessions. The hyperparameter tuned CatBoost models achieved an F1 score for frame-level predictions on a hold-back test set of 0.6670 when trained on player input dimensions only and 0.7725 when game state dimensions were also included. A higher F1 score of 0.9109 was demonstrated when aggregating frame-level predictions to the session level when making an actionable model. When evaluating the session-level errors made by this model, the author noted that the only false positives came from a small set of sessions that were created by the game’s developer. If we exclude the game developer’s own sessions (assuming the developer themselves may have used tools to develop them) the F1 score increases to 0.9583 with a precision of 1.0000. The findings suggest that anti-cheat software can be implemented effectively without the need for direct access to players’ computers and the associated security concerns current anti-cheat software presents. Recommendations for future research include applying the methodology to other games, with specific game recommendations. Expanding the dataset and dimensions used in model training to identify playstyle differences for specific players over multiple sessions. Researching more game state variables that could have predictive value in identifying cheating. As well as collaborating with game developers to improve models and deploy them to game servers.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Weatherton_gwu_0075A_17075.pdf | 2025-04-09 | Open Access |
|