Cross-Institutional Federated Learning System for Secure AI-Driven English Essay Evaluation

Abstract:
Secure and trustworthy essay evaluation is difficult across educational institutions due to privacy concerns, data silos, and variable grading standards. The majority of automated essay scoring systems now in use rely on centralized data aggregation, which raises security concerns, restricts scalability, and compromises scoring equity. This study suggests a Cross-Institutional Federated Learning System based on Secure Federated Transformer Essay Scoring (SF-TES) to address these problems. Decentralized model training is made possible by the framework without requiring the sharing of unprocessed student essays. To enhance discourse comprehension and scoring stability, it combines a hierarchical attention scoring system with a transformer-based semantic encoder. With an accuracy of 0.93, a mean absolute error of 0.64, and an inter-rater consistency of 0.87 as determined by Cohen's kappa, tests performed on the Learning Agency Lab Automated Essay Scoring 2.0 dataset show excellent performance. Exam boards, colleges, and institutions can use the suggested system for large-scale, privacy-preserving essay assessments.