AdaDP-FedSec: adaptive differentially private federated learning with secure aggregation for multi-institutional English learner corpus collaborative training

Abstract

Cross-institutional collaboration in English learner corpus construction promises richer, more representative datasets but faces persistent barriers rooted in data privacy, regulatory compliance, and institutional reluctance to share sensitive learner writing samples. This paper introduces AdaDP-FedSec, a federated learning framework that enables multiple institutions to jointly train corpus-based language models without exchanging raw data. The framework incorporates three integrated mechanisms: an adaptive privacy budget allocation strategy that dynamically calibrates differential privacy noise based on gradient variance and institutional data characteristics, a hybrid secure aggregation protocol combining Shamir secret sharing with Paillier homomorphic encryption to prevent server-side gradient inspection, and a contribution-aware weighted aggregation scheme coupled with a dual-layer personalized model architecture to address cross-institutional data heterogeneity. Experiments conducted across eight simulated institutional nodes on grammatical error detection and writing proficiency classification tasks demonstrate that AdaDP-FedSec recovers roughly three-quarters of the performance gap between standard differentially private federated learning and centralized training, while pushing membership inference attack success close to chance levels. The adaptive budgeting mechanism emerges as the most impactful component, yielding 3–5% point improvements over uniform noise allocation at matched total privacy expenditure. Taken together, these findings point toward a workable—if still early—pathway for privacy-preserving collaborative corpus training in educational NLP.