Technical Summary: Detecting Aimbot Cheaters in MOGs

1. Problem Statement

The rise of Multiplayer Online Games (MOGs) has created a multi-billion dollar industry, but this growth is threatened by cheaters who undermine competitive integrity and player retention. While traditional anti-cheat systems effectively detect memory-based aimbots (which access game process memory to extract player coordinates), they are fundamentally ineffective against a new class of threats: visual aimbots.

Visual aimbots operate entirely outside the game process. They capture the client's screen output and use computer vision models (e.g., YOLO, Faster R-CNN) to detect opponents, subsequently simulating input to aim and shoot. Because they do not interact with game memory, kernel-level anti-cheat solutions cannot detect them. Furthermore, existing disruption-based defenses (such as adding adversarial noise to game frames) often fail because cheaters can iteratively retrain their models to ignore these perturbations or because the perturbations alter the visual design of the game.

2. Methodology: PATCH

The authors propose PATCH, a novel proactive defense strategy that utilizes adversarial patches as honeytokens. Rather than attempting to degrade the cheater's model performance, the goal is to deliberately trigger the attacker's object detection model, causing it to misidentify the patch as a legitimate player.

Core Workflow

1. Patch Generation (Gradient-Based Optimization):

   • The defender generates small, optimized image patches designed to maximize the confidence score of the attacker's object detection model for the "player" class.
   • The process involves initializing a patch with random noise (sampled from $U(0.25, 0.75)$) and iteratively updating pixel values via gradient ascent to increase detection confidence.
   • The optimization keeps the detector's weights frozen and only trains the patch pixels. To ensure location invariance, patches are randomly positioned on training images during optimization.
   • The objective function minimizes the negative mean of the top $k$ class scores ($L = -\frac{1}{k}\sum S_i$), forcing the patch to develop features strongly associated with the target class.

2. Deployment Strategy:

   • HUD Overlay: To preserve the adversarial properties of the patch (which are sensitive to perspective, lighting, and compression), the patches are not rendered as 3D game objects. Instead, they are deployed as Heads-Up-Display (HUD) overlays directly on the client's screen. This ensures the patch appears exactly as trained.
   • Honeytoken Logic:
     • Visual Aimbot + Triggerbot: If the cheater's system aims and fires at the patch, the server receives definitive proof of cheating (shots fired at a non-existent target) and can ban the player.
     • Visual Aimbot Only: If the cheater only aims without firing, the server cannot detect the shot. In this scenario, the server can "flood" the viewport with multiple patches at different locations, causing the cheater's cursor to jitter uncontrollably, rendering the game unplayable for them without affecting honest players.

3. Threat Models:

   • White-Box: The defender has access to the specific model the attacker uses (common as many cheaters purchase pre-trained models).
   • Black-Box: The defender uses a surrogate model (e.g., YOLO11m) to generate patches, which are then tested against attackers using different models (e.g., YOLO12m, YOLO26m) to assess transferability.

3. Experimental Setup

The authors evaluated PATCH using a custom Unreal Engine game and a commercial title (Fortnite).

• Models: The primary base model was YOLO11m (chosen for its balance of speed and accuracy in real-time applications). Transferability was tested against YOLO12m and YOLO26m.
• Configurations:
  • Same-Machine: The aimbot runs on the same PC as the game.
  • External-Machine: The game runs on one PC, screen output is captured via HDMI to a second PC running the aimbot (a common evasion technique).
• Variables: Tests covered 5 patch sizes ($10\times10$ to $30\times30$), 3 screen resolutions ($1440\times1080$, $1920\times1080$, $3440\times1440$), and both white-box and black-box scenarios.

4. Key Results

Detection Effectiveness

• White-Box Performance: In scenarios where the defender knows the attacker's model, PATCH achieved over 90% detection rates (F1-scores) for most patch sizes.
• Black-Box Transferability: When tested against unseen models (YOLO12m and YOLO26m), larger patches ($25\times25$ and $30\times30$) demonstrated 60%–90% cross-model transferability. Even the lower bound of 60% is considered sufficient for reliable identification, as repeated hits on honeytokens provide conclusive evidence.
• Patch Size Trade-off:
  • Smaller patches ($10\times10$) offered better stealth but poor detection rates (approx. 57% in white-box, dropping significantly in black-box).
  • Patches of size $15\times15$ emerged as a critical threshold, achieving 93.4% detection in white-box scenarios while remaining relatively subtle.
  • Larger patches ($30\times30$) provided the highest robustness and transferability but were more noticeable to the human eye.

Robustness and Scalability

• Screen Resolutions: The approach proved scalable across different resolutions. By calculating the scaling factor based on the viewport dimensions relative to the model's training input (640x640), the patches maintained consistent F1-scores across standard, smaller, and ultrawide displays.
• Capture Configurations: Detection rates remained consistent (~92% F1-score) regardless of whether the attacker used a same-machine or external-machine capture setup.
• Fortnite Validation: A proof-of-concept on Fortnite using an open-source visual aimbot (YOLO12) confirmed real-world applicability. The $30\times30$ patches achieved a mean Attack Success Rate (ASR@0.5) of 89.3%, and notably, the patches triggered higher confidence scores than legitimate opponents at high confidence thresholds (0.75).

5. Contributions and Significance

The paper claims the following contributions:

1. Novel Proactive Defense: Introduction of a strategy that uses adversarial patches not to blind the cheater, but to trap them by forcing false positives.
2. Comprehensive Evaluation: Systematic testing of patch parameters (size, location) and execution settings (resolution, capture method) on a custom Unreal Engine environment.
3. Real-World Applicability: Validation of the method on a commercial title (Fortnite), demonstrating that the approach is not limited to custom research environments.

Significance

The authors argue that as visual aimbots become more prevalent, traditional reactive detection is insufficient. PATCH offers game developers a new tool to transform the attacker's own weapon (their object detection model) against them. By deploying honeytokens, defenders can either directly identify cheaters through server-side logs of shots fired at patches or disrupt the cheater's gameplay experience, all while maintaining the integrity of the game for honest players. The paper emphasizes that even with imperfect black-box transferability, the high detection rates of larger patches provide a viable path for mitigating this emerging threat.