A Robust Differential Privacy-Based Fingerprinting Framework for Relational Database Ownership Protection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Protecting privacy and ownership during relational database sharing is essential for secure data-driven analysis and collaboration. However, conventional sanitization and fingerprinting techniques can reduce data utility, alter realvalued attributes, and remain vulnerable to alteration and collusion attacks. This approach employs four real-world datasets obtained from Kaggle—Carbon, Chess, Traffic, and Survive—containing continuous attributes such as Latitude, Longitude, WhiteRatingDiff, BlackRatingDiff, Light, Raining, Age start, and Age end, with an additional identifier retained as the primary key. The preprocessing stage extracts continuous attributes, constructs relational databases, preserves primary keys, and retains label attributes where required for utility evaluation. Unbiased Laplace noise is then injected into nonprimary-key entries using entry-level differential privacy, while group-based fingerprinting and pseudorandom generation support robust detection. Decision Tree classification and regression, Support Vector Machine (SVM) classification, and Gradient Boosted Decision Tree (GBDT) classification and regression are evaluated using prediction accuracy and mean squared error (MSE), while mean absolute error (MAE), fingerprint recovery rate, insertion time, and detection time assess utility, robustness, and efficiency. The proposed scheme achieves a 100% fingerprint recovery rate across alteration and hybrid attack tests and maintains 100% recovery under meancollusion with up to 256 colluders. It also improves utility by up to 20% compared with existing differential-privacy fingerprinting methods.