AUTHOR=Li Long , He Yuanhang , Xu Rui , Chen Bei , Han Boyu , Zhao Yuanyuan , Li Jianhua 

TITLE=Synchronizing LLM-based semantic knowledge bases via secure federated fine-tuning in semantic communication

JOURNAL=Frontiers in Artificial Intelligence

VOLUME=Volume 8 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1690950

DOI=10.3389/frai.2025.1690950

ISSN=2624-8212

ABSTRACT=Semantic communication (SemCom) has seen substantial growth in recent years, largely due to its potential to support future intelligent industries. This advancement hinges on the construction and synchronization of robust semantic knowledge bases (SKBs) across multiple endpoints, which can be achieved through large language models (LLMs). However, existing methods for constructing and synchronizing LLM-based SKBs often face numerous security threats, such as privacy leakage and poisoning attacks, particularly when federated fine-tuning is employed to update LLM knowledge bases. To address these challenges, we propose a novel Secure Federated Fine-Tuning (SecFFT) scheme for synchronizing LLM-based SKBs in semantic communication. First, we incorporate homomorphic encryption into SecFFT to ensure the secure synchronization of model parameters. Second, to enhance the trustworthiness of participants against poisoning attacks, we introduce a residual-based access control mechanism, where only participants with low residuals are authenticated to participate in updating the knowledge base. This mechanism is combined with a hash-based message authentication code. Third, we design a self-adaptive local updating strategy to minimize the impact of poisoned model parameters on benign participants, which is crucial for strengthening the robustness of LLM-based knowledge bases against poisoning attacks. Extensive experiments, conducted using four different datasets from the GLUE benchmark, demonstrate that SecFFT can securely synchronize distributed LLM-based SKBs while maintaining high accuracy (98.4% of the performance of the original federated LoRA), with an acceptable additional cost.