TITLE:
Explainable Privacy-Preserving Scientific Retrieval with Knowledge Graphs: An Evidence-Grounded Framework for Scientific NoSQL Repositories
AUTHORS:
Tanmoy Biswas
KEYWORDS:
Differential Privacy, Explainable Artificial Intelligence, Knowledge Graphs, Large Language Models, NoSQL Databases, Retrieval-Augmented Generation, Scientific Information Retrieval, Trustworthy AI
JOURNAL NAME:
World Journal of Engineering and Technology,
Vol.14 No.4,
October
15,
2026
ABSTRACT: Large Language Models (LLMs) provide powerful interfaces for scientific knowledge retrieval, but privacy-preserving access alone does not satisfy the transparency, provenance, and reproducibility requirements of research-intensive environments. Building directly on prior work that introduced differential-privacy controls for LLM integration with scientific NoSQL repositories [1], this paper proposes Explainable Privacy-Preserving Scientific Retrieval (EPPSR), a unified architecture combining a Scientific Knowledge Graph, semantic and graph retrieval, privacy-risk-aware evidence ranking, provenance, and evidence-level explanations. A revised reproducible V2 benchmark explicitly separates formal differential privacy (DP) from application-layer sensitive-information exposure. Formal DP is evaluated only for a bounded aggregate mean-effect release using replace-one adjacency, clipping to [-18, 18], sensitivity Δ = 6, a Laplace mechanism, an epsilon sweep from 0.1 to 5.0, and repeated-query composition. Sensitive Information Exposure Rate (SIER) separately measures unredacted disclosure of sensitive record identifiers or graph edges. The synthetic benchmark contains 672 records and 80 stratified queries with reliable, contradictory, incomplete-text, sensitive, and lexical-distractor evidence. Across the controlled simulation, EPPSR achieved 100.00% answer accuracy, 97.50% Recall@5, 46.25% citation accuracy, 0.00% SIER, 88.60% provenance completeness, and 92.50% citation support. At epsilon = 1.0, the DP aggregate release achieved 82.02% direction accuracy with a mean absolute error of 6.01 effect-percentage points. Ablation results show that removing consistency reduced answer accuracy to 91.25%, while a semantic-plus-graph-only variant fell to 66.25%. These results demonstrate controlled architectural behavior rather than production LLM performance and show why privacy guarantees, exposure metrics, evidence quality, and retrieval utility must be evaluated separately.