面向联邦学习的学习率裁剪梯度优化隐私保护方案

联邦学习中,攻击者通过模型梯度攻击来恢复训练数据集,使训练数据集的隐私性受到威胁。为保护数据隐私性,差分隐私技术被引入到联邦学习中,但在神经网络训练过程中存在学习率过大导致梯度爆炸不收敛或学习率过小导致梯度收敛过慢的问题,降低学习的准确率。针对上述问题,本文提出一种具有自适应学习率的梯度优化算法(CAdabelief算法),该算法在神经网络中引入学习率裁剪动态界限的概念,动态调整学习率达到理想的值,并趋于稳定;继而将CAdabelief算法引入联邦学习差分隐私框架,提出了面向联邦学习的学习率裁剪梯度优化隐私保护方案,并采用MNIST数据集进行测试实证。实验表明,在相同的隐私预算下,CAdabelief算法训练结果的准确率高于常用的SGD、Adam、Adabelief算法。

裁剪制成率提升方法研究

裁剪制成率作为针织服装企业的一项主要技术经济指标,反映一定质量的面料能够生产出多少件服装,包括排版时面料利用率和裁剪裁成率。为了提升裁剪制成率,文中详细介绍面料利用率和裁剪制成率的含义、影响因素及提高措施。指出可通过增加面料幅宽规格、微调面料幅宽、不同规格尺码模版套排等方式提高面料利用率,使平幅面料的利用率达到75%~80%,圆筒面料的利用率达到80%~85%。通过釆用自动铺布机和自动裁剪机、优化段料长度、严格控制面料幅宽、检测面料克质量、控制环境湿度、适度运用计划加成率,实现动态控制裁剪裁成率,使其稳定在100%~101%,使裁剪制成率不低于面料利用率。

一种面向应用的嵌入式Linux内核混合裁剪方法

针对目前已有嵌入式Linux内核裁剪方法中存在不同程度的粒度偏大问题,提出了一种嵌入式Linux内核混合裁剪方法。根据已有裁剪方法,提出二步裁剪的混合裁剪方法,对RFID手持读写器应用过程中使用到的嵌入式Linux内核进行裁剪。实验表明:混合的裁剪方法降低了嵌入式Linux内核的粒度,裁剪率与其他方法相比提升了24%以上。

Relationship between Living Rate of Bud and Emergence Rate of Ratoon Rice and Characteristics of the First Cropping Mid-season Hybrid Rice

[Objective] The aim was to study the relationship of plant characters with bud living rate and mother stem emergence rate of ratoon rice before harvest. [Method] Eighteen mid-season hybrid rice cultivators approved recently were taken as materials and relationship of plant characters including bud living rate and moth- er stem emergence rate of ratoon rice before harvest was explored, based on relat- ed data. [Result] The shorter rice with lower harvesting index is the main character of high emergence rate for first cropping rice; decline of plant height and increase of ear-bearing percentage would improve emergence of regenerative seedlings. [Conclusion] The research provided scientific reference for breeding of rice cultivars with high regenerative capacity.