A Comprehensive Survey on Robustness and Privacy in Federated Learning Meets Large Language Model at Edge

Large Language Models (LLMs) have revolutionized natural language processing, yet their deployment is hindered by data, computation, and privacy constraints. Federated Learning (FL) offers a promising solution by enabling collaborative, privacy-preserving training across distributed devices, while the push for low-latency on-device intelligence further drives LLM integration into FL and edge settings—posing new challenges in heterogeneity and resource limits. This survey comprehensively reviews the integration of LLMs with federated learning, termed FLM, and its deployment at the edge, with particular emphasis on the robustness, privacy, and trustworthiness challenges that emerge across the LLM lifecycle from pre-training to deployment. We analyze core challenges including communication cost, system heterogeneity, privacy risks, and scalability, with a focus on edge-oriented efficiency techniques such as pruning and quantization. Security vulnerabilities and defenses are also discussed, alongside trade-offs among privacy, robustness, and performance. We further examine demographic, contribution-related, and performance-related biases that can emerge in FLM systems. Finally, we outline open research directions, underscoring the potential of federated and edge intelligence to enable scalable, privacy-aware LLM ecosystems, and aim to offer a unified perspective to guide future research in this fast-moving field.