A Differentially Private Decision Tree-Based Classification Method

As a classical data classification model, the decision tree has been widely applied in behavior prediction and related fields. However, constructing a decision tree without any privacy-preserving mechanism will inevitably risk privacy leakage. Differential privacy, as an effective framework for privacy protection, has been widely adopted in the study of privacy-preserving decision tree construction to protect individual privacy by adding carefully calibrated noise to the count values of the sample data. Existing DP-based solutions commonly suffer from low classification accuracy due to excessive noise injected into redundant count values. To address these issues, we propose a classification gain matrix-based algorithm to systematically store the count values required for information gain computation. Moreover, to reduce the noise level, we utilize the column-sum vector to impose constraints on the count values within the same column of the classification gain matrix. On this basis, we devise a decision tree construction method, DPTBC, which adopts a differential privacy framework to ensure individual privacy. We further introduce a novel order-reconstruction procedure for the classification gain matrix to mitigate the impact of noise on category label count values. The classification gain matrix is then reconstructed according to the column-sum vectors. Theoretical analysis and experimental results demonstrate that DPTBC significantly improves the accuracy of the constructed decision tree while preserving the privacy of the sample data.