Beyond privacy-utility tradeoffs with empirical evidence on privacy-performance gains and fairness independence

Privacy-preserving machine learning is critical if you cannot centralize data or share it with third parties (ethical AI). We propose the first such end-to-end empirical analysis of privacy-utility trade-offs — spanning from model performance, fairness, and explainability across the machine learning pipeline with differential privacy (DP) applied at distinct steps. We analyze nine machine learning models using the User Privacy and Advertising Dataset (UPAD), which includes 50,000 synthetic user records with 29 privacy-aware features, under various ϵ ∈ [0.1, 5.0] privacy budgets. Our results demonstrate that privacy-preserving mechanisms can produce classification accuracy on the order of 54–67%, presenting a direct challenge to the canonical postulate of the privacy-utility trade-off. Importantly, we show that privacy and fairness are orthogonal: both can be satisfied simultaneously for a given algorithm, but achieving strong privacy does not ensure fair outcomes too — on average, violations of demographic parity are 22.2%. We have 80% consensus on feature importance under DP: Interpretability analyses via LIME and SHAP 2.4. A mixed-effects model detects two clusters of performance, with the best possible accuracy ∼67.35% obtainable when ϵ = 0.133. Collectively, these results offer a conceptually grounded and empirically validated framework to go beyond the privacy-utility trade-off with important implications for regulatory compliance and responsible AI deployment.