SQLLM: A Secure and Quantized Framework for Large Language Models in 5G Private Network Operations

Abstract

With the development of 5G technology, the demand for large language models (LLMs) in private network operation and maintenance is growing. These models enhance the intelligence and efficiency of network management through deep learning techniques. However, the application of LLMs for 5G private network operation and maintenance faces the dual challenges of data security and resource limitations. To address these challenges, we propose SQLLM, which is an integrated framework dedicated to detecting abnormal user query-based network attacks and performing high-quality compression on large language detection models for network attacks. Specifically, leveraging the powerful representational capacity of large language models, we utilize the LoRA fine-tuning technique to identify normal questions and three types of attack questions, thus avoiding behaviors that attempt to disrupt or obtain sensitive information. Moreover, this research innovatively adopts the static quantization technique to compress LLMs for private network operation and maintenance. Unlike traditional static quantization that uses fixed parameters and uniform granularity for activations, our optimized strategy introduces a dynamic smoothing factor $\alpha $ to transfer outlier variance from activations to weights and adopts a hybrid per-tensor/per-token granularity tailored to 5G O&M data, thus avoiding the challenge of the sharp decline in the inference ability of the quantized LLM caused by outliers, enabling it to adapt to resource-constrained and multi-user concurrent usage environments. Finally, we validate four types of questions and evaluate the performance of the quantized model on key performance indicators such as accuracy, VRAM consumption, and inference time. The experimental results demonstrate that the SQLLM model achieves excellent performance on all test indicators. This achievement proves the security and feasibility of the efficient deployment of LLMs for private network operation and maintenance in the 5G private network environment with network attacks and resource limitations.