基于微调LLaMA模型的加密货币交易异常检测研究  

Research on anomaly detection in cryptocurrency trading based on fine tuning LLaMA model

作　　者：曾磊 陈志德[1] ZENG Lei;CHEN Zhide(Fujian Normal University,Fuzhou 350117,China)

机构地区：[1]福建师范大学,福州350117

出　　处：《计算机应用文摘》2025年第8期63-65,共3页

摘　　要：加密货币市场的高波动性和监管滞后性显著增加了交易异常检测的复杂性。传统的规则引擎和统计方法难以适应动态的市场环境,而机器学习模型则面临标注数据稀缺和特征工程成本高等双重挑战。为此,基于LLaMAFACTORY框架,提出了一种轻量级的异常检测模型,重点解决低资源场景下的高效适配问题。通过引入LoRA+(低秩适应)技术,利用预训练模型的参数空间特性,在保持泛化能力的同时实现了轻量微调,显著降低了对标注数据和计算资源的依赖。实验结果表明,优化后的LLaMA3.2-3B-instruct模型在标准测试集上达到了97%的检测准确率,且参数更新效率较基线模型提升了32%。The high volatility and regulatory lag in the cryptocurrency market significantly increase the complexity of anomaly detection in transactions.Traditional rule engines and statistical methods are difficult to adapt to dynamic market environments,while machine learning models face dual challenges of scarce annotated data and high feature engineering costs.To this end,a lightweight anomaly detection model based on the LLaMA FACTORY framework is proposed,focusing on solving the efficient adaptation problem in low resource scenarios.By introducing LoRA+(low rank adaptation)technology and utilizing the parameter space characteristics of pre trained models,lightweight fine-tuning is achieved while maintaining generalization ability,significantly reducing dependence on annotated data and computational resources.The experimental results showed that the optimized LLaMA3.2-3B instruction model achieved a detection accuracy of 97% on the standard test set,and the parameter update efficiency was improved by 32% compared to the baseline model.

关 键 词：加密货币 交易异常检测 Lora 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]