Balancing Privacy and Model Performance: An Experimental Analysis of Data Anonymization Techniques

Balancing data privacy with model performance remains a central challenge in deploying machine learning systems on sensitive datasets. This paper presents an experimental analysis of three widely used syntactic data anonymization techniques k -anonymity, l -diversity, and t -closeness and evaluates their impact on both privacy protection and downstream machine learning utility. Using the Adult Census and German Credit benchmark datasets, we systematically measure privacy outcomes through re-identification risk and attribute disclosure risk, and utility outcomes through classification accuracy, AUC, and information loss across multiple learning models. The results demonstrate that$\mathbf{t}$-closeness provides the strongest privacy guarantees, achieving up to 94.2% reduction in re-identification risk, but incurs the highest degradation in model performance. K -anonymity preserves the greatest utility while exhibiting significant vulnerability to attribute disclosure under skewed sensitive attribute distributions. L-diversity consistently offers a balanced tradeoff, maintaining moderate privacy protection with acceptable utility loss. The study further analyzes the influence of dataset characteristics, model choice, and computational overhead on anonymization effectiveness. Based on empirical findings, we derive practical guidelines to support informed selection of anonymization techniques according to application sensitivity and performance requirements, contributing actionable insights for privacy-aware machine learning practice