Privacy-Preserving Cross-Domain Score Fusion for Distributed Public-Service Data: A Federated Architecture and Empirical Benchmark
Journal of Artificial Intelligence General science (JAIGS) ISSN:3006-4023 · 2026
DOI: 10.60087/jaigs.v10i01.493 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Artificial Intelligence General science (JAIGS) ISSN:3006-4023, 2026. Verfasst von Shireesha Mangannagari. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Public-service institutions - healthcare, education, government, and judicial systems - increasingly hold complementary information about the same populations, but regulatory fragmentation and, for sovereign tribal nations, data-sovereignty obligations prevent this information from being combined. This paper presents a federated architecture for privacy-preserving, cross-domain data use across these four domains: per-domain differentially private federated averaging (client-level privacy; Section 3.5), a within-domain multimodal feature-fusion encoder using synthetic image-embedding features (not real medical imaging; Section 3.3), a cross-domain score-fusion stage governed by a pseudonymous record-linkage model, an embedded, sensitivity-tested explainability module, and a CARE-aligned governance layer explicitly distinguished from the architecture's technical privacy guarantees. All results below are computed directly from code execution, with bootstrap confidence intervals and significance tests reported throughout, on a synthetic benchmark (30,500 records; fixed seed) with deterministic data generation and deterministic non-private baselines, and statistically characterized (repeated, CI-reported) stochastic privacy experiments, rather than illustrative figures. The benchmark uses synthetic Gaussian-generated features, a synthetic protected attribute, missing-completely-at-random missingness, and synthetic (non-real) image-embedding features (Section 3.3); results below characterize methodological behavior on this constructed benchmark, not real-world public-service performance. Combining modalities within a domain significantly improved F1 (the harmonic mean of precision and recall at a fixed decision threshold, used here as an operating-point-sensitive complement to AUC's threshold-free ranking measure), but the accompanying AUC change was not statistically distinguishable from no change (paired bootstrap, p=0.798) - the multimodal benefit in this benchmark is F1-specific, not a general performance gain. Federating five simulated clients per domain across 30 rounds changed AUC only slightly in absolute terms, though that small change reached statistical significance in three of four domains (paired bootstrap; Schooling was the exception), while F1 dropped substantially and significantly in three of four domains, most severely in the judicial domain (Government was the exception); a Brier-score and optimal-threshold analysis is consistent with, and supports, an operating-point (threshold-calibration) explanation for this F1 pattern - the F1 gap closes substantially at a re-selected threshold - without conclusively isolating threshold calibration as the sole causal mechanism. Using a paired experimental design (identical underlying noise draws scaled by ε, enabling a paired Wilcoxon test), AUC increased monotonically and mostly significantly with ε in every domain within the tested ε ∈ {0.5, 1, 2, 3, 5} range; a formal client-level privacy budget of ε = 3 was nonetheless the dominant utility cost, reducing AUC by roughly 0.25–0.30 relative to non-private federation. A Bonferroni-corrected empirical outlier test found no domain's demographic-parity gap changed significantly under differential privacy; this should not be interpreted as evidence of equivalence between the non-private and DP fairness outcomes, since the test's statistical power was limited. Cross-domain score fusion significantly outperformed the best single-domain model on matched held-out entities for the experimenter-defined synthetic outcome (AUC 0.732 [95% CI 0.699–0.767] fused vs. 0.691 best single domain, paired-bootstrap p < 0.05), per Section 3.4.1's definition - evidence the fusion mechanism behaves as intended on this benchmark, not a real-world utility claim. A separate sensitivity analysis found fused AUC itself increases monotonically across four tested synthetic correlation strengths, though the fusion advantage over the best single domain does not increase monotonically alongside it (Section 4.6.1) - these are two distinct findings and should not be conflated. Explanation-method agreement (SHAP vs. LIME) was low on the Healthcare model only - the sole domain evaluated for this analysis, not extended to Schooling, Government, or Judicial - (Spearman ρ = 0.04, 95% CI crossing zero), while SHAP's top-ranked features significantly influenced the model's own output under a feature-deletion sensitivity test (Wilcoxon signed-rank p < 1e-14) - indicating explanation-method agreement is a poor proxy for model-output sensitivity (we avoid the term “explanation fidelity”, which implies a formal fidelity criterion this feature-deletion test does not establish). We report a full threat model for the record-linkage mechanism, an explicit definition of the cross-domain outcome variable and its governance, quantify the added privacy cost of protecting fusion-stage risk scores under repeated release, and state explicitly which findings are demonstrated only on a small, five-client synthetic benchmark (a methodological demonstration, not a scalability claim) and should not be read as predictions of real-world or larger-scale operational performance.
Zitieren
APA
Mangannagari, S. (2026). Privacy-Preserving Cross-Domain Score Fusion for Distributed Public-Service Data: A Federated Architecture and Empirical Benchmark. Journal of Artificial Intelligence General science (JAIGS) ISSN:3006-4023. https://doi.org/10.60087/jaigs.v10i01.493
Vancouver
Mangannagari S. Privacy-Preserving Cross-Domain Score Fusion for Distributed Public-Service Data: A Federated Architecture and Empirical Benchmark. Journal of Artificial Intelligence General science (JAIGS) ISSN:3006-4023. 2026. doi:10.60087/jaigs.v10i01.493
BibTeX
@article{Mangannagari2026,
author = {Mangannagari, Shireesha},
title = {Privacy-Preserving Cross-Domain Score Fusion for Distributed Public-Service Data: A Federated Architecture and Empirical Benchmark},
journal = {Journal of Artificial Intelligence General science (JAIGS) ISSN:3006-4023},
year = {2026},
doi = {10.60087/jaigs.v10i01.493}
}
EUVIMED