Osu! Standard Cheat Detector (Transformer-based Classifier)
This repository contains the weights for the osu! standard anti-cheat classifier, trained to detect human-like and blatent cheats in '.osr' replays.
It uses a Cross-Attention Transformer architecture to analyze hit-objects and replay frame kinematics.
🌟 Key Features
- Multi-Cheat Detector: Trained to detect Aim Assist, Aim Correction, Relax and Human Autobot
📊 Model Performance
The 'best.pt' model achived the following performance metrics on the validation dataset:
- Precision: 97.3% to 100.0% (Extremely low to zero false positive rate)
- Recall: 72.0% to 76.0% (Catches the vast majority of subtle aim assists and humanized relax cheats.)
- F1-Score: 82.8%
- Accuracy: 85.0%
🏗️ Model Architecture
The classifier works by matching sliding context windows of beatmap hit-objects with replay cursor frames:
- Window Encoder: Comapre a window of $10$ beatmap objects against a window of $20$ replay frames using self-attention and cross-attention blocks.
- Sequence Aggregator: Aggregates local window embeddings across the entire beatmap.
- Classification Head: Output logit for binary classification (0 = Legit, 1 = Cheated).
🚀 Usage
For inference code, installation instructions, and training scripts, please refer to the GitHub repository:
🏋🏻 Dataset & Training Details
- The used for training and validation was auto-generated with this script
- The config used can be found in this repo.
