Multilingual and Multimodal Federated Small Language Model for Global 6G Edge Intelligence

Abstract

This paper presents a federated learning framework that integrates a compact multilingual small language model with multi- modal data processing capabilities for privacy-preserving edge intelligence in 6G networks. The system uses a 1.3 billion parameter variant of the mT5-small architecture enhanced with cross-modal attention to handle text in over 50 languages, images, and time-series sensor data on resource-limited devices. Federated training incorporates language-aware client selection and differential privacy to ensure balanced representation across linguistic groups and compliance with data protection regulations. A reinforcement learning agent dynamically adjusts training parameters to optimize accuracy, energy consumption, and communication latency under 6G ultra-reliable low-latency communication constraints. Evaluation on multilingual emergency response detection and global health monitoring tasks involving 1200 edge devices across 42 countries demonstrates 95.6% classification accuracy, 94.8% F1-score, 63% reduction in energy usage, and 0.71 ms model update latency. The proposed approach outperforms centralized large models and monolingual federated systems by 3.8 percentage points in accuracy and scales to 15 times more devices while maintaining strong privacy guarantees with a differential privacy bound of ϵ = 0.92. This framework provides a scalable and inclusive solution for next-generation intelligent edge systems in globally connected 6G environments.