SRP-Net: sensitive risk propagation network with asymmetric cross-attention for educational data

Abstract

Abstract In the digital transformation of education, detecting sensitive information in structured educational data faces the challenge of semantic mismatch between titles and content, as well as isolated risk assessment that ignores field relevance. We propose the Sensitive Risk Propagation Network (SRP-Net), which is a sequential two-stage architecture that aligns within the field and then propagates across fields. The first stage adopts asymmetric cross attention, treating field names as queries to align with instance values and statistical fingerprints (key values). In the second stage, these aligned representations are input into a sensitivity aware graph neural network through batch dynamic subgraph extraction to avoid excessive smoothing. We have developed Edu KG BERT for domain adaptation through structured knowledge injection and developed the open-source Edu-Sens, which is the first benchmark dataset for education sensitive data synthesized using privacy preserving protocols. SRP-Net achieved an accuracy of 0.8779 Macro-F1 and 95.58% accuracy on Edu-Sens, which is 3.06% higher than the non graph baseline and significantly better than traditional machine learning and pure semantic baselines. The ablation study confirmed the synergistic necessity of explicit in field alignment and dynamic cross field propagation.