OUT-Guard: Context-Aware Risk Detection and Adaptive De-Identification of Rare and Extreme Information

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).

Records containing rare and extreme information can pose substantial disclosure risks even after direct identifiers have been removed. Existing methods typically assess numerical outliers, rare categories, small contextual cells, and behavioral anomalies independently, limiting their ability to detect records made sensitive by overlapping risk factors and to select appropriate de-identification treatments. This paper proposes OUT-Guard, a context-aware framework that integrates numerical-tail, rare-category, contextual-combination, and behavioral-shift risks at the record level and adaptively applies risk-specific de-identification. OUT-Guard prioritizes records with multiple active risk types, generalizes categorical and contextual attributes first, reassesses residual risk, and modifies numerical values only when tail risk persists within a multi-risk pattern. The framework was evaluated on 1,007,913 transactions from Online Retail II and 2,044,737 personal records from the 2024 American Community Survey Public Use Microdata Sample. Multi-risk records represented 0.316% and 0.307% of the respective datasets. Default OUT-Guard reduced multi-risk records by 81.1% and 97.6%, respectively, while retaining 100% of the selected aggregate numerical utility measures. The enhanced version increased risk reduction to 94.4% and 99.4%, with utility retention of 96.6% and 99.97%, respectively. Sensitivity, ablation, and scalability analyses demonstrated stable behavior and practical scalability to million-record datasets. OUT-Guard supports privacy-preserving sharing of sensitive structured data.