Implementasi K-Anonymity Untuk Mengurangi Risiko Identifikasi Pada Dataset Data Pribadi

Sharing personal data for research and analysis creates a risk of identifying individuals, because removing direct identifiers alone is not enough. This study implements k-anonymity on 16,007 records of the UCI Adult dataset using five quasi-identifiers (age, sex, race, marital status, native country), with income as the sensitive attribute. The method is full-domain generalization with record suppression (5% limit), evaluating all 180 combinations of generalization levels for k = 2, 3, 5, and 10 in Python. The results show that unique records fell from 1,352 (8.45%) to zero for all values of k, maximum re-identification risk fell to 1/k, and average risk fell from 0.1442 to 0.0341 (k = 2) and down to 0.0045 (k = 10). This protection came with a rise in total information loss from 0.0922 to 0.2471, with 4.2% to 4.7% of records suppressed and age fully hidden at k = 10. In this configuration, k = 5 was a balanced choice between privacy and utility. However, 1.51% to 12.65% of records still sit in classes that are homogeneous in income, leaving them vulnerable to homogeneity attacks, so k-anonymity should be complemented with l-diversity or t-closeness.