Style-Conditioned Privacy-Preserving Synthesis of Student Writing

Building NLP applications for education is constrained by data scarcity due to student privacy concerns and data regulations. To address this, we propose a style-aware synthesis approach for reconstructing student writing to retain analytical value while reducing identifiability. Our method isolates content, structure, and style as separate conditioning channels and aggregates over student-disjoint exemplar partitions within style clusters, extending Differentially Private In-context Learning (DP-ICL) with cluster-restricted retrieval. Across three educational NLP datasets, our approach achieves high style fidelity, preserves downstream task performance, and reduces authorship attribution risk across style-based and semantic attacks.