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

Abstract

Abstract Cloud-edge Internet of Things (IoT) continuums increasingly rely on artificial intelligence to detect intrusions and enforce zero-trust policy at scale, yet purely neural detectors remain difficult to audit and purely symbolic engines struggle to generalize across noisy, heterogeneous telemetry. Building on a neurosymbolic zero-trust orchestration framework proposed for United States federal cloud-edge IoT continuums, this study develops and evaluates a generalized neuro-symbolic architecture for trustworthy predictive analytics that couples a neural pattern-recognition component with an explicit, auditable symbolic rule layer. Three formal research questions are addressed: whether a symbolic-only rule engine is significantly less accurate than a neural-only classifier (RQ1), whether a confidence-gated neuro-symbolic hybrid preserves neural-level accuracy (RQ2), and what the quantified, statistically grounded missed-detection cost of the hybrid's symbolic override mechanism is (RQ3). Using a disclosed, literature-informed synthetic dataset of 12,000 cloud-edge IoT telemetry records spanning five traffic classes (benign, DDoS, botnet, MITM/spoofing, and reconnaissance), three configurations a neural-only Random Forest classifier, a symbolic-only zero-trust rule engine, and a confidence-gated neuro-symbolic hybrid 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. The symbolic-only engine was significantly less accurate than the neural-only classifier across all 30 trials (paired t-test, p < .0001; Cohen's d = -59.9), confirming H1. The neuro-symbolic hybrid showed a statistically significant but practically small accuracy reduction relative to neural-only (mean difference -0.32 percentage points, paired t-test p < .0001, Cohen's d = -1.72), a real, disclosed effect rather than the exact tie suggested by the single illustrative split, addressing RQ2 with the nuance repeated resampling reveals. The hybrid's missed-detection rate was significantly higher than neural-only's (mean increase 0.72 percentage points, p < .0001, Cohen's d = 4.68), confirming H3 and quantifying the trade-off beyond a single-split estimate, while its explainability coverage rose from 0% to 100% of decisions by architectural design. The study contributes a generalizable neuro-symbolic reference architecture, a reproducible evaluation and statistical-validation methodology for the accuracy-explainability trade-off in cloud-edge IoT security analytics, and transparent, hypothesis-tested empirical evidence including the honest disclosure that small, consistent effects can be statistically significant without being practically large. Limitations, policy implications for regulated critical-infrastructure deployments, and directions for future empirical validation on real network traffic are discussed.