Abstract
This paper introduces an unsupervised machine learning framework to detect collusion—unfair cooperation between opposing teams—in large-scale multiplayer games like Battle Royales. By combining social network analysis with in-game behavioral features and the Isolation Forest algorithm, the researchers successfully identified high-probability colluders in real-world datasets without requiring massive sets of pre-labeled "cheater" data.
Introduction
In competitive gaming, "hard" cheating like aimbots is relatively easy to flag because it involves "super-human" actions. However, collusion is a "soft" form of cheating where two teams who are supposed to be enemies secretly work together to eliminate others.
Collusion is incredibly difficult to detect for two reasons:
- It looks "human": Colluders don't use forbidden software; they coordinate via external chats (like Discord) to avoid fighting each other.
- Scale: With millions of players and hundreds of thousands of matches, game designers cannot manually review every game. Furthermore, there is very little "labeled data" (confirmed cases of collusion) to train traditional supervised AI models.
Methodology
The authors propose a hybrid approach that blends human domain expertise with unsupervised learning.
Feature Engineering
Instead of looking for "super-human" stats, the framework looks for "teammate-like" behavior between opponents. Five key features were extracted:
- Landing Proximity: How close do the two "opposing" teams land at the start of a match?
- Rank Placement Difference: Do they consistently finish the game with nearly identical ranks (e.g., 1st and 2nd)?
- Acquaintance: Have these "opponents" played as teammates in the past?
- Match Frequency: How often do they end up in the same match together?
- Consecutive Matches: How often do they appear in back-to-back games?
Social Network Analysis and Isolation Forest
The researchers used graph theory to visualize player relationships, identifying "clusters" of players who are suspiciously intertwined. Since collusion is rare, it is treated as an anomaly. The authors used an Isolation Forest (IF)—an unsupervised algorithm that "isolates" outliers by randomly partitioning data. The fewer partitions it takes to isolate a data point, the more likely it is to be a colluder.
Experimental Setup
The study utilized two large-scale, real-world datasets from a team-based tournament mode:
- Dataset 1: 113,060 unique matches and 173,603 unique players over 3 days.
- Dataset 2: 44,097 unique matches and 128,342 unique players over 2 days. This dataset included a small number of "confirmed" colluders previously identified by human experts to help validate the model.
Key Insights: The "Teammate in Disguise" Hypothesis
The core insight is that colluders essentially act like a single large team. The probability of two random teams landing near each other and finishing 1st and 2nd across multiple matches is statistically near zero. When this happens, it is a strong indicator of premeditated coordination.
As illustrated in the figure above, the distance between spawn locations for colluders (Dataset 2, bottom-left) is significantly lower than for typical opponents, mirroring the behavior of actual teammates. Similarly, the rank differences for these suspicious pairs remain consistently low even as the number of matches played increases.
Results
The model proved highly effective at identifying suspicious behavior:
- High Precision: Out of the top 20 most "abnormal" pairs identified by the algorithm in Dataset 1, game designers confirmed that 14 were most likely colluding. In Dataset 2, 16 out of the top 20 were confirmed.
- Efficiency: Manual review of a single suspicious pair takes a designer about 3 minutes. By filtering the population down to just a few hundred outliers, the system saves hundreds of hours of manual labor.
- Validation: The system successfully flagged all previously "confirmed" colluders that were hidden in the test datasets.
Conclusion and Future Work
The authors emphasize that the AI acts as a "force multiplier" for game designers, narrowing down millions of players to a small "watchlist" for manual review. Future work aims to improve temporal resolution by accounting for mid-game behavior, expanding dataset duration to detect subtle long-term colluders, and moving toward real-time detection while matches are still in progress.
Summary updated