PFDF: Privacy Preserving Federated Decision Forest for Classification

Tongyaqi Li, Qingqiang Qi, Chengyu Hu, Xuelei Li, Peng Tang, Shanqing Guo

Abstract: The rapid development of data analysis technology and easily accessible datasets enable the construction of comprehensive analysis models, promoting decision-making processes in various fields. Decision trees, known for their high interpretability and cost-effectiveness, has become a common choice for decision-making in areas like housing price prediction and medical forecasting. However, in federated learning, there exists a risk of individual privacy leakage. To address this challenge, we propose a new privacy-preserving decision tree boosting model (PFDF). To protect data holders’ privacy, we adopt differential privacy technology to perturb sensitive data that might lead to privacy leakage. Under the premise of privacy preservation, this model includes a novel approach for designing the global attribute selection and leaf node judgment scheme, considering the data imbalance among data holders. Additionally, continuous splitting values are generated using a clustering method. To enhance classification accuracy, our “multi-perspective” decision tree boosting scheme considers the optimal attributions of previously constructed trees. Accuracy tests on several benchmark datasets demonstrate that our scheme outperforms classical and the state-of-the-art approaches currently known to us in both centralized and federated learning.