基于模型剪枝的深度神经网络分级授权方法的实现
Implementation of a Hierarchical Authorization Method for Deep Neural Networks Based on Model Pruning
宋允飞1
作者信息
- 1. 贵州师范大学 大数据与计算机科学学院,贵州 贵阳 550025
- 折叠
摘要
基于现有深度神经网络模型无法根据使用权限进行分级授权的问题,设计提出了一种新型的DNN模型分级授权方法,其可以根据模型权限不同分发不同模型精度.该方法依据模型剪枝技术实现了模型性能的分级,利用特定的剪枝速率或剪枝阈值对模型进行剪枝和微调,通过在剪枝和微调阶段对模型进行调整从而使模型输出不同等级的准确率,最后将不同的用户权限与对应等级的准确率相匹配.在多个数据集和DNN模型上进行了实验,并利用CIFAR-10和CIFAR-100数据集进行验证.实验结果表明,该方法能够有效地将模型的性能分级,在多个DNN模型上都有良好的效果.
Abstract
Based on the problem that existing Deep Neural Networks models cannot be hierarchically authorized according to the usage authority,a novel hierarchical authorization method for DNN models is designed and proposed,which can distribute different model accuracies according to different model permissions.The method realizes model performance grading based on the model pruning technique,which uses a specific pruning rate or pruning threshold to prune and fine-tune the model.By adjusting the model during the pruning and fine-tuning stages,the model outputs different levels of accuracy.Finally different user privileges are matched with the corresponding level of accuracy.Experiments are conducted on multiple datasets and DNN models and validated by using CIFAR-10 and CIFAR-100 datasets.The experimental results show that the method is effective in grading the performance of the model and works well on several DNN models.
关键词
深度神经网络/分级授权/版权保护/模型剪枝Key words
Deep Neural Networks/hierarchical authorization/copyright protection/model pruning引用本文复制引用
基金项目
中央引导地方科技发展专项(QKZYD[2022]4054)
贵州省教育厅自然科学研究项目(QJJ[2023]011)
贵州师范大学学术新秀基金(QSXM[2022]31)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据