TLDR: A research paper introduces AntiCheatPT, a transformer-based machine learning model, and CS2CD, a public dataset, to detect cheating in Counter-Strike 2 using gameplay data. The model achieved 89.17% accuracy and 93.36% AUC, emphasizing reproducibility and offering a robust, open-source solution for server-side cheat detection.
Cheating in online competitive computer games, such as Counter-Strike 2 (CS2), significantly undermines the experience for honest players. Traditional anti-cheat systems like Valve Anti-Cheat (VAC) struggle to keep pace with new cheating methods without becoming overly intrusive. This challenge is compounded by the lack of publicly available data and open-source research in the field of cheat detection.
Addressing this critical gap, a new research paper introduces AntiCheatPT 256, a transformer-based machine learning model designed to identify cheating behavior in Counter-Strike 2. Alongside the model, the researchers have publicly released CS2CD, a comprehensive, labeled dataset of 795 CS2 matches, making a significant contribution to academic research in this area.
The CS2CD Dataset: A Foundation for Research
The creation of a robust dataset was a cornerstone of this research. The team chose Counter-Strike 2 due to its prevalent cheating issues and the availability of match replays (DEM files), which contain detailed records of player actions. Data was collected from csstats.gg, a match sharing service that indicates if a player has been VAC-banned. This information was used to identify potential cheaters.
To ensure accuracy, all 317 matches containing at least one VAC-banned player were manually reviewed. Players were labeled as cheaters only when there was undeniable evidence, prioritizing the minimization of false positives. The raw DEM files, which contain sensitive information, underwent an anonymization process before publication. The final CS2CD dataset, weighing in at 48.9 GB, is available on Hugging Face, providing a valuable resource for future studies.
AntiCheatPT 256: A Transformer in Action
The AntiCheatPT 256 model leverages a transformer encoder architecture, inspired by the groundbreaking “Attention Is All You Need” paper. For the model’s input, gameplay data from the CS2CD dataset was segmented into “context windows.” Each window, spanning 256 ticks (equivalent to 4 seconds) and centered around a player kill, captures 44 data points describing the attacker, victim, weapon, and map information.
To enhance the model’s robustness and prevent overfitting to specific in-game locations, data augmentation was applied. Gaussian noise was added to player coordinates, ensuring the model focused on relative positioning rather than absolute locations. Importantly, cheater data was augmented more heavily to balance the dataset, resulting in a 1:2 ratio of cheater to non-cheater context windows.
Also Read:
- The Fair Game: A Dynamic Approach to Ensuring AI Fairness Over Time
- New Insights into Multi-Winner Voting Through Data Analysis
Performance and Reproducibility
The AntiCheatPT 256 model was trained using a 70/15/15 split for training, validation, and testing, ensuring that no context windows from the same match appeared in different sets. On an unaugmented test set, the model achieved an accuracy of 89.17% and an Area Under the Curve (AUC) of 93.36%. These results are comparable to or exceed those of other limited studies in behavioral cheat detection for Counter-Strike games.
A key emphasis of this research is reproducibility. Unlike previous works that often lack architectural details or data access, the AntiCheatPT project provides all code and model weights as open-source. This transparency allows other researchers to build upon this work, fostering further advancements in the field.
The paper concludes that an open-source, reproducible method for behavioral cheat detection is feasible for game companies, especially those that prefer to avoid kernel-level anti-cheats. This approach demonstrates that even hardware-level cheats can potentially be detected through server-side machine learning systems. For more details, you can read the full research paper here.
