Abstract
In competitive multiplayer games, collusion—often referred to as "teaming"—occurs when two or more opposing teams collaborate to gain an unfair advantage. Detecting these behaviors is a significant challenge for game designers due to the massive scale of player populations and the subtle nature of the offense. This paper presents an unsupervised framework that identifies colluding players by analyzing social relationships and in-game behavioral patterns. By utilizing graph theory and Isolation Forests, the system measures the degree of collusion between opposing pairs and highlights high-probability outliers for manual review. The approach was validated on two real-world datasets, each containing over 170,000 unique players and 100,000 matches, demonstrating high precision and operational efficiency.
Introduction
In the ecosystem of Battle Royale games, such as Apex Legends or Warzone, fair play is maintained through the detection of traditional cheats like aimbots or wallhacks. These "blunt" cheats involve superhuman actions that are relatively easy to flag via automated telemetry. However, collusion presents a more nuanced problem. Collusion occurs when teams that are intended to be enemies decide to share resources, avoid combat with one another, or coordinate attacks against other unsuspecting teams.
Because colluding behavior often mimics legitimate gameplay—just directed at the wrong targets—it is difficult to distinguish from normal player interactions. Furthermore, the sheer volume of matches makes manual oversight impossible, and the scarcity of confirmed cases prevents the use of traditional supervised machine learning models. This research addresses these hurdles by proposing an unsupervised anomaly detection system designed to act as a decision-support tool for game security teams.
Methodology
The proposed detection framework relies on a hybrid approach combining domain-specific feature engineering with unsupervised learning.
Feature Engineering
To capture the "signature" of collusion, the researchers defined a feature set based on the psychology and logistics of teaming. Key features include:
- Proximity: The physical distance between opposing teams throughout a match. Colluders tend to stay close to one another without engaging in combat.
- Rank Difference: The similarity in final match placement. Colluding teams often survive together until the final circles, resulting in nearly identical end-of-match rankings.
- Acquaintance: A social metric determining if the "enemies" have previously played as teammates in other matches.
- Frequency: The regularity with which specific pairs of opponents appear in the same match, which may indicate "match sniping" or intentional queue coordination.
Isolation Forest (IF)
Given the lack of labeled training data, the system utilizes the Isolation Forest algorithm. This technique is specialized for anomaly detection by isolating observations through random partitioning. In this context, because colluders are rare and their behavior deviates significantly from the standard player base, they are isolated much faster than normal players, resulting in a higher anomaly score.
Social Network Analysis (SNA)
Beyond individual match behavior, the system employs graph theory to visualize player relationships. By representing players as nodes and their interactions as edges, the researchers can identify suspicious clusters where supposedly opposing players exhibit tighter social and behavioral connections than the general population.
Datasets and Validation
The study utilized two large-scale datasets from a popular team-based multiplayer game:
- Dataset 1: 113,060 matches and 173,603 unique players collected over three days.
- Dataset 2: 44,097 matches and 128,342 unique players over two days. This dataset included a small "ground truth" set of confirmed colluders previously identified by human moderators.
Results and Discussion
The model proved highly effective at narrowing the search space for game designers. In Dataset 2, the Isolation Forest successfully flagged 100% of the previously confirmed colluders as high-scoring anomalies.
To test the system's precision on unknown data, game designers performed a manual review of the top 20 most abnormal pairs identified by the model. The results were compelling:
- In Dataset 1, 14 out of the top 20 pairs (70%) were confirmed as colluders.
- In Dataset 2, 16 out of the top 20 pairs (80%) were confirmed as colluders.
Beyond accuracy, the system provides significant operational efficiency. Manual review of a single case typically takes approximately three minutes. By filtering millions of potential pairings down to a few hundred high-probability candidates, the framework makes manual enforcement feasible at a global scale.
Conclusion
This research demonstrates that while collusion is difficult to detect through raw telemetry alone, the combination of social network analysis and unsupervised anomaly detection provides a robust solution. The study highlights that colluders essentially leave a "teammate signature" in their movement and match outcomes.
Future work will focus on temporal dynamics—analyzing how the movement of colluding teams evolves in real-time throughout a match—and expanding the analysis over longer time windows to identify serial offenders who may attempt to mask their behavior across different sessions. Ultimately, this system serves as a vital tool in maintaining the competitive integrity of large-scale multiplayer environments.
Summary updated