Federated Multimodal Transformers: Enabling Secure and Collaborative Learning Across Edge–Cloud Environments

Abstract

The rapid proliferation of edge devices, Internet of Things (IoT) applications, and multimodal data sources has created unprecedented opportunities for intelligent decision-making while simultaneously raising significant concerns regarding data privacy, communication efficiency, and computational scalability. Conventional centralized deep learning frameworks require the transfer of sensitive data to cloud servers, making them vulnerable to privacy breaches and regulatory constraints. This research presents a Federated Multimodal Transformer (FMT) framework that enables secure, privacy-preserving, and collaborative learning across heterogeneous edge–cloud environments without exposing raw data. The proposed architecture integrates transformer-based multimodal feature extraction with federated learning to jointly process diverse data modalities, including images, text, audio, and sensor streams distributed across geographically dispersed edge devices. Secure aggregation mechanisms, adaptive federated optimization, and edge-aware communication strategies are incorporated to reduce bandwidth consumption while preserving model performance. Furthermore, the framework supports heterogeneous device capabilities through dynamic client selection and personalized model adaptation, thereby improving scalability and robustness in real-world deployments. Experimental evaluation demonstrates that the proposed FMT framework achieves superior multimodal representation learning, enhanced predictive accuracy, reduced communication overhead, and stronger privacy guarantees compared to conventional centralized and federated approaches. The proposed methodology offers a practical solution for privacy-sensitive domains such as healthcare, smart cities, autonomous transportation, industrial IoT, and intelligent surveillance, where collaborative intelligence is essential but direct data sharing is restricted. The findings highlight the potential of federated multimodal transformers as a next-generation paradigm for secure, scalable, and intelligent distributed learning in edge–cloud ecosystems.