Stacking Ensemble Defense Framework for Large Model‐based English Proficiency Assessment in Wireless Educational Networks

Wireless educational networks increasingly deploy large language models for automated English proficiency assessment, enabling ubiquitous learning across mobile and distributed environments. However, these LM‐based assessment systems remain vulnerable to adversarial manipulation, where learners subtly alter input features to inflate scores without genuine improvement. This vulnerability undermines assessment integrity and pedagogical value. We propose a stacking ensemble defense for English proficiency assessment (SEDEF‐EPA) defense framework employing a stacking ensemble architecture. SEDEF‐EPA comprises an adversarial detection layer with three heterogeneous LM assessors (transformer‐encoder, sequence‐to‐sequence, and acoustic‐linguistic model) and a multilayer perceptron detector analyzing prediction distributions. Through adversarial training of the detector only, SEDEF‐EPA maintains computational efficiency. A joint decision mechanism prevents over‐reliance on individual assessors. Experiments on the automated student assessment prize dataset demonstrate over 94% accuracy in non‐adversarial scenarios and exceed 87% under white‐box attacks.