Comparative Evaluation of Privacy-Preserving Techniques for Real-Time Big Data Analytics: An Empirical Study and Hybrid Differential Privacy–Synthetic Data Framework

Big Data analytics creates substantial value but also increases the possibility of identity disclosure, linkage attacks, and misuse of sensitive records. This study analytically compares five privacy-preserving techniques—Random Noise Addition (RNA), Differential Privacy (DP), Data Swapping (DS), K-Anonymity (KA), and Synthetic Data Generation (SDG)—under a common experimental framework designed for real-time Big Data analytics. The experiments use a standardized synthetic dataset containing 500 individual profiles and 15 demographic, financial, and behavioural attributes. Six numerical variables are normalized, categorical variables are encoded, and privacy resistance is tested with a Re-identification Neural Network (RNNR). Privacy, utility, latency, throughput, scalability, distributed-platform execution time, robustness, and regulatory suitability are jointly evaluated. The results show that DP with the Laplace mechanism at ε=1.0 preserves numerical means with 99.97–100.00% accuracy and produces 52% RNNR re-identification accuracy, close to the 50% chance baseline. Spark execution is 0.6 s, compared with 3.4 s on Hadoop. SDG using a Gaussian Copula provides the strongest structural privacy and high statistical fidelity but requires 8.4 s on Spark. KA performs poorly for streaming use: k=5 requires 48.6 ms per record and remains vulnerable to background-knowledge attacks. A multi-criteria evaluation ranks DP highest among individual techniques (62/70), followed by SDG (57/70), while the proposed Hybrid DP+SDG architecture reaches 67/70. The findings indicate that privacy-preserving analytics should be selected by deployment context rather than by a single universal criterion.