基于联邦学习的火灾图像检测算法  

Fire image detection algorithm based on federated learning

作　　者：杨帆[1] 吴浩宇 郭雅婷 董迪昊 YANG Fan;WU Hao-yu;GUO Ya-ting;DONG Di-hao(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi′an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021

出　　处：《陕西科技大学学报》2022年第6期184-191,共8页Journal of Shaanxi University of Science & Technology

基　　金：陕西省社会发展科技攻关计划项目(2016SF-418)。

摘　　要：针对火灾图像检测背景下传统机器学习缺乏数据集,模型泛化性弱的问题,提出一种基于联邦学习的火灾图像检测算法.通过联邦学习收集分散的边缘火灾图像数据,解决数据集单一问题,通过多方协同训练综合模型,提升模型的收敛效率与泛化性.为了降低联邦学习中的加密开销,使用基于中国剩余定理门限秘密共享安全方案替代同态加密方案.实验结果表示,与中心化学习相比,前5轮收敛速率提升至少一倍,在30轮全局迭代后,准确率可以达到90.21%,正确率达到88.72%,误报率低至3.41%.实验所用加密方案所需时间开销约为训练总时间1%,远低于使用同态加密的联邦学习方案.Aiming at the problem that traditional machine learning lacks data set and model generalization is weak in the background of fire image detection,a fire image detection algorithm based on federated learning is proposed.The scattered edge fire image data is collected through federated learning to solve the single problem of data set.The neural network is used to improve the convergence efficiency and generalization of the model through multi-party collaborative training of the comprehensive model.In order to reduce the encryption cost in federated learning,a secret sharing security scheme based on the Chinese residual theorem threshold is used to replace the homomorphic encryption scheme.Experimental results show that compared with the centralized learning,the convergence rate of the first 5 rounds increases at least twice.After 30 rounds of global iteration,the accuracy rate can reach 90.21%,the accuracy rate can reach 88.72%,and the false positive rate is as low as 3.41%.The time cost of the encryption scheme used in the experiment is about 1%of the total training time,which is much lower than that of the federated learning scheme using homomorphic encryption.

关 键 词：图像检测 联邦学习 模型泛化 数字信封 同态加密 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]