AI-Driven Personalized Decision Support for Secure and Adaptive Digital Services

Digital service providers increasingly seek to personalize user experience through artificial intelligence, yet personalization depends on collecting fine-grained behavioral and, in some cases, accessibility-related data whose exposure raises privacy and security concerns, particularly for users who disclose support needs to receive better-adapted service. Building on a study demonstrating that AI-driven, data-informed personalization improves adaptive learning, communication, and therapeutic accessibility on a digital health platform for students with autism spectrum disorder, this study generalizes that personalization paradigm into a secure, adaptive personalization architecture applicable across digital service domains, incorporating an explicit, tunable privacy safeguard. Three formal research questions are addressed: whether personalization significantly improves accuracy and accessibility-need detection relative to a non-personalized baseline (RQ1), whether adding differential privacy significantly increases the accessibility-need miss rate relative to unprotected personalization (RQ2), and what the quantified, statistically grounded accuracy cost of that privacy protection is (RQ3). Using a literature-informed synthetic dataset of 10,000 user-interaction records spanning four engagement outcomes, three configurations, a non-personalized baseline, a fully-personalized AI model, and a privacy-preserving personalized AI model using differentially-private noise on sensitive features were evaluated on an illustrative held-out split and, to test the study's hypotheses rigorously, across 30 independently resampled train/test trials with paired statistical testing and a fixed privacy budget. The fully-personalized configuration was significantly more accurate and had a significantly lower accessibility-need miss rate than the non-personalized baseline across all 30 trials (accuracy: mean 71.7% versus 44.6%, p < .0001, Cohen's d = 29.7; accessibility-need miss rate: mean 6.2% versus 46.3%, p < .0001, Cohen's d = 14.7), confirming H1. The privacy-preserving configuration showed a statistically significant, quantified accessibility-need miss-rate increase relative to fully-personalized (mean 7.2% versus 6.2%, mean difference 1.09 percentage points, p < .0001, Cohen's d = 1.29), confirming H2, and a significant, small accuracy cost (mean 71.0% versus 71.7%, mean difference 0.68 percentage points, p < .0001, Cohen's d = 1.58), confirming H3. Unlike the single illustrative split, which suggested privacy protection preserved accessibility benefit almost without cost, the rigorous 30-trial validation confirms both costs are real, consistent, and statistically distinguishable from zero, while remaining small relative to the large gains personalization delivers over the non-personalized baseline. The study contributes a generalized, reusable secure personalization architecture, a reproducible privacy-budget tuning and statistical-validation methodology, and transparent, hypothesis-tested empirical evidence quantifying both the substantial benefit of personalization and the smaller, but real and disclosed, equity cost of adding privacy protection to it. Limitations, implications, and directions for validation on real platform data are discussed.