Federated Ensemble Distillation for Robust Text Processing in Multimodal Knowledge Graph Construction

We propose Federated Ensemble Distillation, a novel framework for robust text processing in multimodal knowledge graph (KG) construction that addresses low-resource Named Entity Recognition (NER), modality heterogeneity, and privacy constraints. Our method integrates federated adapter-based ensemble training with cross-modal knowledge distillation, where lightweight client-specific adapters are distilled into a global model while dynamically weighting contributions based on data quality. A differential privacy-preserving consensus voting system resolves entity ambiguities across modalities. Experiments on low-resource NER benchmarks demonstrate a 12.3% F1-score improvement over traditional federated approaches using RoBERTa-large and CLIP-ViT. Our work establishes a new direction for privacy-sensitive multimodal knowledge graph construction in data-scarce environments.