Privacy Protection in the Era of Big Data and Artificial Intelligence

Abstract

The rapid growth of Big Data and Artificial Intelligence has significantly increased the collection and analysis of personal data, raising important privacy concerns in-cluding re-identification, membership inference, and model inversion attacks. Single-technique approaches such as anonymisation or Differential Privacy applied indepen-dently offer limited protection against adversaries possessing auxiliary background knowledge. This thesis investigates a hybrid approach combining k-Anonymity and Differential Privacy (ε-DP) on the UCI Adult Income dataset, where k-Anonymity first reduces re-identification risks through quasi-identifier generalisation and equiva-lence class formation, followed by Differential Privacy mechanisms based on Laplace noise and Randomised Response with varying privacy budgets. A LinearSVC classifier is trained and evaluated using accuracy, precision, recall, and F1-score. k-Anonymity (k = 5) achieves F1 = 64.20% versus 53.58% for the unprotected base-line, confirmed by 5-fold cross-validation, and attributable to feature space compression and implicit regularisation rather than data leakage. Differential Privacy alone sub-stantially reduces model utility at strong privacy budgets, while the hybrid approach outperforms it: at ε = 1.0, the hybrid reaches F1 = 58.33% compared to 32.43% for Differential Privacy on raw data, suggesting that k-Anonymity preprocessing reduces noise sensitivity and improves model stability. The study highlights the potential benefits of combining structural anonymisation with statistical privacy mechanisms in privacy-preserving machine learning, while discussing the limitations of the proposed approach and the trade-off between privacy protection and model utility.