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

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Semantic communication (SemCom) is crucial for intelligent industries, relying on synchronized semantic knowledge bases (SKBs).
• Large language models (LLMs) are key to constructing these SKBs, but federated fine-tuning introduces security vulnerabilities like privacy leakage and poisoning attacks.

Purpose of the Study:

• To propose a novel Secure Federated Fine-Tuning (SecFFT) scheme for synchronizing LLM-based SKBs in SemCom.
• To address security threats including privacy leakage and poisoning attacks during federated updates.

Main Methods:

• Incorporation of homomorphic encryption for secure parameter synchronization.
• Implementation of a residual-based access control mechanism with hash-based message authentication code to counter poisoning attacks.
• Design of a self-adaptive local updating strategy to mitigate the impact of poisoned parameters on benign participants.

Main Results:

• SecFFT securely synchronizes distributed LLM-based SKBs.
• The scheme maintains high accuracy, achieving 98.4% of the performance of federated LoRA.
• Experimental validation on four GLUE benchmark datasets confirms effectiveness.

Conclusions:

• SecFFT offers a robust solution for secure SKB synchronization in SemCom.
• The proposed methods enhance trustworthiness and resilience against adversarial attacks.
• The scheme provides a secure and efficient approach to updating LLM knowledge bases in distributed environments.