XGuardian Anti-Cheat

Elimination-scope analysis Match-scope analysis

Note

1. HUDs (in-game UI overlays, e.g., health bar, navigation map, avatar's ID, kill info, etc.) that may disclose any information are hidden;

2. X-Ray (highlighted character borders can be seen through obstacles) is activated for better illustration purpose. In real-world matches, the normal players will not obtain this function.

3. The legends of the explainable trajectories are idnetical both in the video and the interactive images below.

Default display.

Tick

Feature

Elimination

This page demonstrates how XGuardian generates the explainable final match prediction. First, it produces an original prediction for each elimination, as shown in the table below.

Elimination #	Original Prediction After the Aggregation
1	0.684117
2	0.6946051
3	0.6873236
4	0.7693209
5	0.73794997
6	0.7006488
7	0.6901594
8	0.6728578
9	0.7000603
10	0.726555
11	0.6647711
12	0.67824334
13	0.6937227
14	0.6991382
15	0.68439555
16	0.6942821
17	0.70512176
18	0.6243448
19	0.6593147
20	0.6644799
21	0.70384914
22	0.68810034
23	0.69289875
24	0.6848471
25	0.5973644
26	0.52587
27	0.60854065
28	0.63035697
29	0.663132
30	0.53972197
Average (Input of the Match Prediction)	0.6722031113333333

The mean-aggregated original predictions serve as input to the random forest model for match prediction. Through its decision-making process, the model generates XGuardian's final judgment. The following diagram illustrates how the sample player in the above table was determined. In this diagram, X-axis indicates the input value of the sample for the given feature; Y-axis denotes the adjusted SHAP values. For more intuitive visualization, the original SHAP values (computed by the model interpreter) are offset so that the value corresponding to the dynamic threshold probability equals 0. This value reflects the difference between the sample's predicted positive-class probability and the dynamic threshold. A transformed SHAP value of 0 indicates the prediction probability exactly matches the dynamically computed threshold on the training set, while positive/negative values represent deviations above/below the threshold; All non-case-study samples are colored dimgray; Circles represent ground-truth cheaters; Triangles represent ground-truth normal players; The case study sample is highlighted with a purple star marker. The sample's input value of 0.6722031113333333 lies above the decision boundary, causing the model to classify it as positive (i.e., a cheating player).

XGuardian: Towards Generalized, Explainable and More Effective Server-side Anti-cheat in First-Person Shooter Games

Explainability Case Study