FedCMB: A Cross-Modal Federated Learning Framework for Multi-Modal Cyberbullying Detection with Shared Feature Space

The increasing multimodality of cyberbullying poses significant challenges to detection tasks, making traditional single-modal approaches inadequate. Federated learning has emerged as an effective distributed learning paradigm for collaborative training while preserving user privacy, but existing methods generally overlook the heterogeneity of client models and modality discrepancies. To address these issues, we propose FedCMB, a cross-modal federated learning framework tailored specifically for multi-modal cyberbullying detection. FedCMB constructs a modality-invariant shared feature space to effectively fuse features from heterogeneous client models. Additionally, we introduce a knowledge distillation-based aggregation method to facilitate efficient knowledge transfer and integration from heterogeneous client models to a unified server-side model. Our experiments show that FedCMB enhances accuracy by 11.4% over baselines while preserving privacy.