Federated Large Language Models for Privacy-Preserving Edge Intelligence: Architectures, Challenges, and Performance Optimization

Sushmita Singh, Mukul Garg, Parul Verma, Nitin Jain, Ayushi Singh

Abstract


The rapid convergence of Large Language Models (LLMs) and Edge Computing has established Edge Intelligence, enabling real-time, context aware, and personalized natural language understanding directly on client end-devices. However, deploying multi-billion-parameter foundation models on resource-constrained hardware introduces severe computational, VRAM memory, and network bandwidth bottlenecks. Furthermore, centralizing sensitive user text on third-party cloud servers raises profound data privacy risks under regulatory frameworks such as GDPR and HIPAA. Federated Learning (FL) mitigates these privacy risks by facilitating collaborative model adaptation across decentralized edge nodes without centralizing raw user data. This review paper presents a comprehensive survey on Federated Large Language Models (FedLLMs) for privacy-preserving edge intelligence. We categorize system architectures (centralized server-orchestrated, peer-to-peer decentralized, and hierarchical edge-cloud topologies), Parameter-Efficient Fine-Tuning (PEFT) integrations including Federated Low-Rank Adaptation (FedLoRA) and 4-bit Quantized FedQ-LoRA, and privacy mechanisms such as Differential Privacy (DP) and Secure Aggregation. In addition, we benchmark system-level performance regarding communication cost, memory footprint, and downstream language perplexity, outlining key research gaps and strategic future directions.

KEYWORDS: Federated Learning, Large Language Models, Edge Intelligence, Differential Privacy, Parameter-Efficient Fine-Tuning, Low-Rank Adaptation, On-Device AI, Communication Efficiency.

Full Text:

PDF 108-120

Refbacks

  • There are currently no refbacks.


Copyright 2017 ManTech Publications. All Rights Reserved