Jianwei Yang, Zehua Cheng · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840649
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
The scaling of Large Language Models (LLMs) heavily relies on high-quality proprietary data, yet strict privacy regulations (e.g., GDPR) render vast amounts of real-world text containing Personally Identifiable Information (PII) inaccessible. Existing sanitization techniques struggle with implicit reference leakage and contextual degradation. In this paper, we propose a paradigm grounded in the hypothesis that LLM reasoning capabilities stem predominantly from the invariant logical context rather than from specific PII entities. We introduce a Graph-Driven Context-Preserving Anonymization Framework comprising three phases: (1) Global Graph Construction to explicitly model dependencies, (2) Targeted Topological Knock-out to isolate and prune explicit subject nodes and implicit coreference edges, and (3) Contextual Re-alignment for autoregressive reconstruction. Across the proprietary CCC-500 corpus and the public Text Anonymization Benchmark (TAB), evaluated over three random seeds against 18 heuristic, transformer-based, differentially private, and LLM-driven baselines, our method reaches 100.0±0.00% Deterministic Desensitization Rate on CCC-500 and 99.8±0.10% on TAB under parser-completeness conditions, while preserving downstream Top-1 Accuracy within 0.1 percentage points of unsanitized text on identity-orthogonal multi-hop QA. An isolation ablation confirms that both the graph topology and the constrained regenerator contribute independently to the result. All improvements over the strongest generative baseline (GPT-4o Few-shot) are significant at p<0.001 on a paired Wilcoxon signed-rank test.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.