model_card.md

Model Card: XGBoost Exploit Pattern Classifier

Model Details

• Architecture: Tree-Based Ensemble (XGBoost)
• Strategy: One-vs-Rest (OvR) Multi-Label Classification
• Framework: scikit-learn & xgboost
• Hyperparameters:
  • n_estimators: 300
  • max_depth: 3
  • learning_rate: 0.1
  • subsample: 0.8

Intended Use

• Primary Use Case: Real-time classification of EVM transaction execution traces into 10 known exploit categories.
• Out-of-Scope: Detection of non-EVM patterns, or zero-day patterns not present in the training set.

Training Data

• Dataset Size: 120+ labeled exploit samples.
• Data Split: 80% Training, 20% Testing (Stratified).

Evaluation Results

• Macro F1 Score: 0.1833
• Target Achieved: No (Goal >= 0.80)

Caveats and Limitations

• The model depends heavily on the accuracy of the Trace Feature Extractor.
• Requires maintenance and retraining as new DeFi protocols emerge.