Resource Management and Security Challenges for Deploying and Adapting Large Language Models in Fog Computing

Authors:

Abstract
Fog computing is an emerging distributed paradigm that extends cloud services to the network edge, enabling low-latency, context-aware processing for Internet of Things (IoT) applications. By decentralizing computation and storage to fog nodes near end devices, fog architectures significantly enhance real-time responsiveness compared to centralized clouds. As large language models (LLMs) gain prominence in AI-driven services, their deployment in fog environments introduces critical challenges due to their intensive computational and memory requirements. This paper provides a comprehensive analysis of resource management and security challenges associated with deploying and adapting LLMs in fog computing. We review state-of-the-art techniques in task scheduling, load balancing, and virtualization, and examine how these are impacted by the unique heterogeneity and resource constraints of fog infrastructures. Particular attention is given to model partitioning strategies, communication-computation tradeoffs, and adaptive deployment across diverse fog nodes. In addition, we explore key security risks posed by distributed LLM deployment, including data leakage, model theft, gradient inversion, and denial-of-service attacks. We evaluate emerging countermeasures such as differential privacy, homomorphic encryption, encrypted inference, and robust training techniques, assessing their suitability for fog-based learning frameworks. Finally, we highlight open research directions, including dynamic model partitioning, energy-efficient inference, security certification protocols, and the development of standardized benchmarks for evaluating LLM performance and privacy in fog environments. Our findings aim to support researchers and practitioners in addressing the trade-offs between efficiency, scalability, and security in deploying large-scale AI models at the network edge.

This publication has 16 references indexed in Scilit: