Researchers introduced a privacy-preserving hybrid federated learning framework to detect deepfakes without sharing sensitive datasets. The approach, FedHybrid-ViT, combines convolutional and transformer components across heterogeneous data sources. The system reported an AUC of 0.925, indicating strong classification performance under federated constraints. For organizations subject to data residency rules or with fragmented data holdings, this is positioned as a way to improve detection models without aggregating raw content. The work also adds to the broader push for on-device or distributed learning methods that can reduce compliance barriers in AI deployment.