CrowdStrike Researchers Develop Custom XGBoost Objective to Improve ML Model Release Stability

CrowdStrike researchers present a custom XGBoost objective that perturbs gradients and Hessians to preserve decision-value ranking across model releases, improving stability and reducing surprise false positives in a PE malware classifier. Across two experiments (119k and 285k samples), swap-in false positives decreased by ~18.75% and ~10.2% respectively, with virtually no change in true positives, and the report includes the mathematical basis, an intuitive example, and Python implementation guidance.