Application and legal response of ensemble learning based on hyperparameter optimization in privacy protection

To address the issues of weak model generalization, high risk of privacy leakage and lack of legal compliance in network intrusion detection, this paper proposes an ensemble learning protection framework (HO-EL-PP) that combines hyperparameter optimization and differential privacy. The framework uses a tree-structured Parzen estimator (TPE) to jointly optimize and stack integrate XGBoost, LightGBM, and CatBoost, and injects Laplacian noise into the meta-learner output to achieve differential privacy, while combining SHAP values to build a compliance audit mechanism. Experiments on CIC-IDS2017 (2.83 million samples) and IoT-23 datasets show that HO-EL-PP achieves an accuracy of 99.21% and an F1 value of 98.98% on the CIC-IDS2017 test set, and a single inference takes only 2.3 ms. Even in the high privacy interval ( ε = 0.1), the model accuracy rate remains above 86%, and the comprehensive compliance score is as high as 0.875, effectively achieving an excellent balance between technical utility and legal compliance.