面向网络中心运维的轻量化安全大模型微调方法OA

The intelligent upgrade of network-centric operation and maintenance systems relies on the efficient deployment of large language models(LLMs).However,resource-constrained edge environments and dynamic adversarial threats pose dual challenges to traditional fine-tuning methods.Existing approaches often address computational efficiency and model security in isolation,leading to lightweight strategies that compromise defense capabilities or security mechanisms that exacerbate re-source consumption.To resolve this conflict,this paper proposes a Lightweight and Secure Co-Tuning Framework(LST-Framework)that achieves multi-objective optimization through dynamic parameter sparsification and hierarchical defense mechanisms.The method first constructs a task-sensitive dynamic parameter activation strategy,identifying redundant weights via spectral analysis of the Hessian matrix to establish sparse fine-tuning paths,reducing computational complexity to 23%of full-parameter fine-tuning.Simultaneously,a three-tier defense system—comprising adversarial input detection,lightweight gradient masking,and output uncertainty verification—enhances robustness against white-box attacks with only an 8%increase in inference overhead.Experiments on typical network operation datasets(e.g.,CIC-IDS2017)demonstrate that the LST framework maintains 98.3%original task accuracy while compressing memory usage by 51.7%,reducing infer-ence latency by 32.4%,and successfully resisting 94.6%of PGD-targeted attacks.This work provides theoretical and tech-nical foundations for deploying large models in low-resource,high-security scenarios such as 5G core networks and industrial IoT.