The new era of decentralized, privacy-oriented social media platforms has brought us a set of related enforcement problems which include detecting cyberbullying, disinformation on a coordinated scale5,14. These centralized or unimodal systems are unable to work efficiently when faced with stringent privacy concerns or multimodal content. In this paper, we present Federated Cross-Modal Graph Transformer (FCMGT) to jointly model text, image and audio features and social graph structure in federated learning settings. Furthermore, the proposed approach is enhanced by a dynamic adversarial training to mitigate content perturbation, graph manipulation and model-poisoning attacks. On a large-scale synthetic decentralized dataset (2 M + interactions), the experiments reveal that FCMGT achieves an F1-Score of 0.927, outperforming the best baseline by 4.6%, and achieves an AUC of 0.963. Performance drop down under adversarial attacks is only 3.8%, in contrast to 15–30% for previous models. These findings position FCMGT as a reliable, scalable, and privacy-preserving system for safe guarding next-generation decentralized social networks.
