A Multi-Dimensional Empirical Evaluation of k-Anonymity, Centralized and Local Differential Privacy, and Federated Learning for Personal Data Protection in Big Data Scenarios

Abstract

This study systematically evaluates the performance of mainstream personal information anonymization technologies in the big data environment. It focuses on k-anonymity family models, differential privacy (DP) and its local variants, and privacy protection mechanisms in federated learning (FL). A multi-dimensional comprehensive evaluation system is constructed in the study, measured from four aspects: privacy protection strength, data utility, computational efficiency, and scalability. Indicators such as normalized certainty penalty (NCP), area under the curve (AUC), and processing time are adopted. The experimental results show that under the condition of medium-intensity privacy protection (privacy budget ε = 1.0), the central differential privacy (CDP) mechanism exhibits the best comprehensive performance in most scenarios. In the medical data release scenario, the NCP of CDP is 0.05, the model's AUC is 0.945, and the average processing time is 25 seconds. In contrast, the k-anonymity model has higher information loss (NCP = 0.22) and lower model accuracy (AUC = 0.920). In the FL scenario of financial anti-fraud, the AUC of the global model under CDP protection is 0.890, which is significantly better than the AUC of the local differential privacy (LDP) scheme (0.842), with comparable computational overhead. In the user behavior analysis experiment involving one million users, the LDP method based on the Randomized Aggregate Privacy-Preserving Ordinal Response (RAPPOR) mechanism achieves lower mean absolute error and higher popular category recognition accuracy (99.0%). In all tests, the NCP and AUC of CDP remain stable with changes in data scale, while the processing time increases approximately linearly with data volume, showing excellent scalability. The above results quantitatively verify the significant advantages of CDP over traditional k-anonymity and LDP methods in data utility, privacy protection strength, and computational efficiency. Thus, empirical evidence can be provided for the selection of anonymization technologies in large-scale data processing scenarios.