Adaptive Hierarchical Evidence Fusion for Sensitive Field Detection in Structured Data: A Gated Residual Correction Network

Automatic detection of sensitive fields in structured data is a critical prerequisite for privacy compliance and data governance. However, existing approaches face severe cross-domain generalization challenges. Hand-crafted pattern rules often fail under highly heterogeneous naming conventions, while single statistical models tend to overfit and degrade sharply under distribution shifts between training and deployment domains. These limitations stem from the weak semantic signals and distributional heterogeneity of structured data, which make it difficult to simultaneously capture explicit rules and latent, variant-sensitive attributes. To address these challenges, we propose a detection framework based on multi-view complementary features and a Hierarchical Gated Residual Network (HGRN). The framework first constructs a full-spectrum feature system that integrates explicit rules and implicit statistical fingerprints (e.g., entropy and character texture) to fill the semantic gap. It then introduces a decision mechanism combining robust priors with dynamic residual calibration: a random forest provides a stable probabilistic anchor, which is further nonlinearly corrected by a learnable gating-and-expert network. This design explicitly resolves the cognitive conflict between rule-dominated regions and complex distributional regions. Experiments on multiple real-world datasets—including DeSSI, CMS Open Payments and Home Credit—show that the proposed method achieves a Macro-F1 of 0.9408 on DeSSI and exhibits strong in-domain performance. Under strict frozen-model cross-domain transfer, HGRN mitigates the catastrophic collapse observed in pure neural baselines and maintains moderate detection capability, offering interpretable trust allocation between rule-based priors and data-driven correction in both financial and healthcare scenarios.