A Robust Approach for Multi Classification-Based Intrusion Detection through Stacking Deep Learning Models

Intrusion detection is a predominant task that monitors and protects the network infrastructure.Therefore,many datasets have been published and investigated by researchers to analyze and understand the problem of intrusion prediction and detection.In particular,the Network Security Laboratory-Knowledge Discovery in Databases(NSL-KDD)is an extensively used benchmark dataset for evaluating intrusion detection systems(IDSs)as it incorporates various network traffic attacks.It is worth mentioning that a large number of studies have tackled the problem of intrusion detection using machine learning models,but the performance of these models often decreases when evaluated on new attacks.This has led to the utilization of deep learning techniques,which have showcased significant potential for processing large datasets and therefore improving detection accuracy.For that reason,this paper focuses on the role of stacking deep learning models,including convolution neural network(CNN)and deep neural network(DNN)for improving the intrusion detection rate of the NSL-KDD dataset.Each base model is trained on the NSL-KDD dataset to extract significant features.Once the base models have been trained,the stacking process proceeds to the second stage,where a simple meta-model has been trained on the predictions generated from the proposed base models.The combination of the predictions allows the meta-model to distinguish different classes of attacks and increase the detection rate.Our experimental evaluations using the NSL-KDD dataset have shown the efficacy of stacking deep learning models for intrusion detection.The performance of the ensemble of base models,combined with the meta-model,exceeds the performance of individual models.Our stacking model has attained an accuracy of 99%and an average F1-score of 93%for the multi-classification scenario.Besides,the training time of the proposed ensemble model is lower than the training time of benchmark techniques,demonstrating its efficiency and robustness.

基于Stacking模型的学术论文多标签分类系统构建

学术论文高质量多标签自动分类是推动学术研究发展的关键程序之一。本研究利用Stacking模型将随机森林、支持向量机、极限树、极端梯度提升和神经网络五个分类器融合为一个异质集成分类器,并利用基于问题转换思想的多二分类模型将该分类器应用于学术论文多标签分类。根据学术论文的特点,依次实现了与之配套的论文特征提取模块、TF-IDF加权模块、数据预处理模块,最终构建成一个面向学术论文的多标签分类系统。仿真实验验证了本研究构建的学术论文多标签分类系统在处理学术论文多标签分类问题时,较传统的单模型分类器或同质集成模型分类器在泛化能力、稳定性与准确率方面都有一定程度的提升。图9。参考文献21。

Crustal thickness beneath the Tanlu fault zone and its tectonic significance based on two-layer H-κstacking

Tanlu fault zone(TLFZ)is the largest active fault zone in eastern China.It is characterized by complex tectonic evolution and multiple faults and marks the boundary between the North and South China blocks.An indepth understanding of the distinct crustal structures of both parts of the TLFZ will provide valuable insights into the lithospheric and crustal thinning in eastern China,extensive magmatism since the Mesozoic,and formation mechanisms of metallogenic belts along the Yangtze River.In this study,a two-layer H-κstacking approach was adopted to estimate the thicknesses of the sediment and crystalline crust as well as the corresponding vP/vS ratios based on high-quality teleseismic P-wave receiver functions recorded by permanent and temporary stations in and around the TLFZ.The geological units in the study region were delineated,especially the crustal structures beneath extensive sedimentary basins on both sides of the TLFZ.The following conclusions can be drawn:(1)The crustal thickness in and around the TLFZ greatly varies depending on the segment.In the northern segment,the crust is relatively thin beneath the eastern part of the Songliao Basin,a broad uplift of the Moho can be observed,and the Moho descends from south to north.The crust below the central and southern segments becomes thinner from west to east.The thickness of the crust is less than 30 km toward the eastern side of the boundary between the Jiangsu and Anhui provinces,that is,significantly thinner than in other areas.In terms of the vP/vS ratios,high anomalies were detected in the central-southern segments of the TLFZ,indicating the upwelling of deep mantle magma via deep faults.(2)Positive isostatic gravity anomalies were observed in the eastern part of the northern segment of the TLFZ and in the eastern part of the Suwan segment.The crustal thickness is smaller than that obtained from the Airy model of isostasy.This suggests that the lower crust in this area may have experienced intensive transformation processes,which may be related to crustal thinning(caused by crustal extension)and the strong uplift of the mantle in eastern China.The isostatic gravity anomalies between the eastern and western parts of the TLFZ indicate that the fault zone plays a dominant role in controlling the development of the deep crustal structure.(3)Significant crustal thinning was observed beneath the eastern part of the boundary between the Jiangsu and Anhui provinces in the southern segment of the TLFZ,suggesting that this area is prone to lithospheric thinning of the North China Craton.Due to the subduction,compression,and retreat of the Paleo-Pacific Plate during the Yanshanian Period as well as the dehydration of subducting oceanic crust(within subduction zones),the asthenosphere and oceanic crust in eastern China partially melted,resulting in mantle enrichment.The basic magma from the mantle is accumulated at the base of the crust,leading to magmatic underplating.In areas with weak topography toward the east of the TLFZ,magma rises to the upper crust and surface,resulting in the enrichment of multiple metal deposits in this area.

VMD-Stacking集成学习的多特征变量短期负荷预测模型

针对当前短期电力负荷预测结果准确度不够高的问题,提出一种由变分模态分解(variational modal decomposition, VMD)和Stacking集成学习框架组合的多特征变量短期负荷预测模型。在预测前使用VMD算法将负荷数据分解,然后加入对模型重要性较高的特征变量,再建立由轻量级梯度提升机(light gradient boosting machine, LightGBM)与极限梯度提升机(extreme gradient boosting, XGBoost)融合的Stacking集成学习预测模型,并比较不同天气情况下对预测模型准确度的影响。经实际算例对比验证表明:多特征的VMD-Stacking集成学习预测模型的误差较小。采用VMD算法分解历史负荷序列,分解后子模态分量的周期性体现了出来,让模型预测波动性较大的负荷时更容易;温度、天气、农历和节假日情况等影响负荷变化的关键因素有被考虑到,模型的准确度得以提高;Stacking集成学习模型对各算法取长补短,泛化能力增强,预测的准确度高于单一模型。

基于熵权法Stacking集成学习的多分类窃电检测

为了精准定位窃电行为,减小电力窃取给电力系统带来的经济损失,提出了一种基于熵权法Stacking(stacking based entropy,E_Stacking)集成学习的多分类窃电检测模型。首先基于用电量信息共线性的特点,使用方差膨胀因子(variance inflation factor,VIF)作为标准对数据降维,以降低数据复杂度。然后在模型训练时嵌入k折交叉验证,有效防止模型过拟合。该模型包含初级学习器和元学习器两层学习器,可以充分结合两层学习器的优点,将学习的互补特征和判别特征相结合,进一步提高检测性能。最后,使用爱尔兰数据集和部分加州大学欧文分校(University of California Irvine,UCI)数据集验证模型,结果优于目前几种常见的方法,证明该模型的有效性和稳定性。

Design Optimization of CFRP Stacking Sequence Using a Multi-Island Genetic Algorithms Under Low-velocity Impact Loads

A method to improve the low-velocity impact performance of composite laminate is proposed, and a multi-island genetic algorithm is used for the optimization of composite laminate stacking sequence under low-velocity impact loads based on a 2D dynamic impact finite element analysis. Low-velocity impact tests and compression-after impact（CAI） tests have been conducted to verify the effectiveness of optimization method. Experimental results show that the impact damage areas of the optimized laminate have been reduced by 42.1% compared to the baseline specimen, and the residual compression strength has been increased by 10.79%, from baseline specimen 156.97 MPa to optimized 173.91 MPa. The tests result shows that optimization method can effectively enhance the impact performances of the laminate.

基于Stacking多模型融合的储能电站电池预制舱消防预警

由于储能电站电池预制舱监测温度由多个传感装置独立组成,导致对高温异常进行消防预警的灵敏度偏低,为此,提出基于Stacking多模型融合的储能电站电池预制舱消防预警方法研究。引入了Stacking集成学习方法,对多元储能电站电池预制舱温度数据进行融合处理,并利用正则系数对融合后数据的偏差进行校正;在消防预警阶段,结合温度参数表现的发展趋势以及消防预警的时间尺度要求,设置了个性化的预警温度,根据融合结果与预警温度之间的关系,判断是否作出预警处理。在测试结果中,设计方法对于不同程度异常温度均实现了有效的预警,具有良好的灵敏度。

Modification of the Hybridization Gap by Twisted Stacking of Quintuple Layers in a Three-Dimensional Topological Insulator Thin Film

Twisting the stacking of layered materials leads to rich new physics. A three-dimensional topological insulator film hosts two-dimensional gapless Dirac electrons on top and bottom surfaces, which, when the film is below some critical thickness, will hybridize and open a gap in the surface state structure. The hybridization gap can be tuned by various parameters such as film thickness and inversion symmetry, according to the literature. The three-dimensional strong topological insulator Bi(Sb)Se(Te) family has layered structures composed of quintuple layers(QLs) stacked together by van der Waals interaction. Here we successfully grow twistedly stacked Sb_2Te_3 QLs and investigate the effect of twist angels on the hybridization gaps below the thickness limit. It is found that the hybridization gap can be tuned for films of three QLs, which may lead to quantum spin Hall states.Signatures of gap-closing are found in 3-QL films. The successful in situ application of this approach opens a new route to search for exotic physics in topological insulators.

基于多模型融合Stacking集成学习方式的负荷预测方法

人工智能及机器学习技术的快速发展,为负荷预测问题提供了崭新的解决思路。该文结合人工智能的前沿理论研究,提出一种基于多模型融合 Stacking 集成学习方式的负荷预测方法。考虑不同算法的数据观测与训练原理差异,充分发挥各个模型优势,构建多个机器学习算法嵌入的 Stacking集成学习的负荷预测模型,模型的基学习器包含 XGBoost树集成算法和长短记忆网络算法。算例使用 ENTSO 中瑞士负荷数据对算法有效性进行了验证。预测结果表明,XGBoost、梯度决策树、随机森林模型能够通过自身模型的增益情况对输入数据的特征贡献度进行量化分析;Stacking中各个基学习器的学习能力越强,关联程度越低,模型预测效果越好;与传统单模型预测相比,基于多模型融合的Stacking 集成学习方式的负荷预测方法有着较高的预测精度。

基于Stacking集成学习的夏玉米覆盖度估测模型研究

以基于无人机多光谱影像提取的夏玉米植被指数作为特征变量,利用皮尔森相关系数结合随机森林反向验证权重的方法进行特征选择,去除冗余特征。以随机森林、梯度提升树、支持向量机和岭回归作为初级学习器,以岭回归作为次级学习器,建立基于Stacking集成学习的夏玉米覆盖度估测模型,并通过5折交叉验证进一步提升模型泛化能力,采用随机搜索和网格搜索结合的方法对模型超参数进行优化,使用4种回归指标进行模型精度评价,并利用次年数据验证其鲁棒性。结果表明,与单一模型以及决策树、Xgboost、Adaboost、Bagging集成框架相比,Stacking集成学习模型具有更高的精度和更强的鲁棒性,R^(2)为0.9509,比单一模型平均提升0.0369,比其他集成模型平均提升0.0417;Stacking集成学习模型RMSE、MAE和MAPE分别为0.0432、0.0330和5.01%,各指标分别比单一模型平均降低0.0138、0.0130和2.14个百分点,分别比其他集成模型平均降低0.0185、0.0126和2.15个百分点。本研究为夏玉米覆盖度估测提供了新的方法。

基于Stacking与多特征融合的加密恶意流量检测

加密技术保护网络通信安全的同时,大量恶意软件也采用加密协议来隐藏其恶意行为。在现有基于机器学习的TLS加密恶意流量检测模型中,存在单模型检测算法对多粒度特征适用性差和混合流量检测误报率高的问题。提出基于Stacking策略和多特征融合的非解密TLS加密恶意流量检测方法。分析加密恶意流量特征多粒度的特点,提取流量的流特征、连接特征和TLS握手特征。对所提取的特征通过特征工程进行规约处理,从而减少计算开销。对规约处理后的3类特征分别建立随机森林、XGBoost和高斯朴素贝叶斯分类器模型学习隐藏在流量内部的规律。在此基础上,使用流指纹融合处理后的多维特征,利用Stacking策略组合3个分类器,构成DMMFC检测模型来识别网络中的TLS加密恶意流量。基于CTU-13公开数据集对构建的模型进行性能评估,实验结果表明,该方法在二分类实验上识别召回率高达99.93%,恶意流量检测的误报率低于0.10%,能够有效检测非解密的TLS加密恶意流量。

基于多特征和Stacking算法的Android恶意软件检测方法

Android由于其广泛的普及率使得其平台上的恶意软件数量不断增加,针对目前大部分方法采用单一特征和单一算法进行检验,准确率不高的不足,提出了一种基于多特征与Stacking算法的静态检测方法,该方法能够弥补这两方面的不足.首先使用多种特征信息组成特征向量,并且使用Stacking集成学习算法组合Logistic,SVM,k近邻和CART决策树多个基本算法,再通过训练样本进行学习形成分类器.实验结果表明,相对于使用单一特征和单一算法其识别准确率得到提高,可达94.05%,该分类器对测试样本拥有较好的识别性能.

基于Stacking多算法融合模型的数据资源定价方法训练流程如下

[目的/意义]针对现有数据资源定价标准模糊、交易规则缺失、买卖双方信息不对称的问题,提出基于Stacking多算法融合模型的数据资源定价方法。[方法/过程]首先,基于四分位法及Box-Cox方法对样本数据进行预处理,剔除异常数据的同时调整数据分布,并基于Lasso算法以均方根误差最小为目标完成特征筛选;其次,基于Pearson系数选取最优算法组合,构建基于Stacking集成学习融合多种机器学习算法的数据资源价格预测模型;最后,以国信优易数据平台交易数据为样本进行算例分析。[结果/结论]相较于单算法模型及其他集成算法模型,XGBoost、SVR、KNN、MLP融合模型的预测精度及性能均明显提升;实现了数据资源自身和市场的双向反馈平衡,形成闭环的价值链,为数据资源定价提供参考。

基于XGBoost与Stacking融合模型的恶意程序多分类检测方法

当前在恶意程序多分类检测领域,传统静态和动态检测方法受反取证技术影响较大;在新型基于网络流量的检测方法中,由于各类恶意程序流量特征的相似性较大,使用人工提取的数据流特征和传统机器学习方法不能取得较高的准确率。针对上述问题,文章提出一种基于XGBoost与Stacking融合模型的恶意程序多分类检测方法。在获取目标恶意程序对外通信流量并自动提取初始网络特征后,对初始数据集进行预处理和多重特征选择,而后使用基于XGBoost的特征创造算法,在初始特征基础上自动化生成高级特征集,并结合Stacking集成算法实现多模型融合以提升恶意程序多分类检测的准确率。在此过程中,为减少寻找最优参数组合的时间,使用贝叶斯优化方法确定各个模型的最优参数组合,并采取多种正则化策略解决模型过拟合问题。实验结果表明,与其他传统方法相比,该检测方法在恶意程序多分类的准确率上有较大提升。

基于多模型融合Stacking集成学习的油田产量预测

基于机器学习前沿理论,提出一种基于多模型融合Stacking集成学习方式的组合预测方法,以国内某特高含水油田区块中多口水驱产油井历年生产历史数据为试验样本,预测其动态产油量。依据不同算法的训练原理,选取极限梯度提升树算法、长短记忆网络(LSTM)、时域卷积网络(TCN)等作为模型的基学习器,采用多元线性回归作为模型的元学习器。结果表明:融合后的Stacking模型充分发挥了各基学习器的优势,相比单一模型,融合后的Stacking模型预测平均误差较小,预测鲁棒性较好。该模型的提出对融合模型在特高含水油藏开发方面具有重要的应用意义。

Engineering Study on the Treatment of Multi-soil-layering System on the Rural Domestic Wastewater in Taihu Basin

By combining with the actual situation in the rural area,the practical technology of domestic wastewater treatment which had the wide popularization value was developed in the rural area of Taihu Basin.Moreover,the multi-soil-layering system was used to treat the concentrated rural domestic wastewater,and the demonstration project was established in Fenshui Village,Yixing,Jiangsu.The result showed that the infrastructure and operating cost of system was low,and the treatment effect was good.The average removal ratios of COD,NH+4-N,TN,TP and SS were respectively 70%,83%,59%,76% and 94%.The quality of yielding water could reach Grade A standard of Pollutant Emission Standards in Urban Wastewater Treatment Plant.

基于多模型融合Stacking集成学习的异常用电检测方法研究

针对单一异常用电检测方法对于存在不平衡性的数据集检测效率普遍不高的问题,提出了一种基于多模型融合Stacking集成学习的异常用电检测方法。首先,以居民用电数据作为研究对象,分析用户在习惯上表现的不同特征,结合不平衡处理技术和分类预测算法进行研究;其次,为了提高模型的整体性能,采用量子遗传算法对集成学习模型中的参数做优化处理;最后,通过云南某地区用电数据集进行验证,证明所提模型相比单一学习模型检测的准确率有明显提升,对提升异常排查效率,降低电力公司的运营成本具有重要意义。

基于多因子融合和Stacking集成学习的大坝变形组合预测模型

变形是反映大坝服役形态变化的直观表征,构建高效准确的变形预测模型对于大坝结构安全控制十分重要。传统单因子及单算法变形预测模型存在泛化能力不足、鲁棒性差等问题,易出现预测偏差甚至误判。针对这一问题,本文选取不同变形解释因子及回归算法,构建多种单因子单算法预测模型,结合Stacking集成学习思想,对上述模型进行组合,提出了大坝变形组合预测模型。该组合模型以Stacking集成学习为核心,采用高斯过程回归作为元学习器,从算法、因子两方面对单因子单算法预测模型进行集成,并通过k折交叉验证减小模型过拟合风险。以某混凝土拱坝变形监测数据为例,通过多模型构建与性能比较,对所提出模型的准确性与有效性进行评估。结果表明:单因子单算法预测模型具备准确性和多样性的特征;通过算法、因子集成,组合模型的预测精度和鲁棒性得到了显著提高,在水位波动期的预测能力得到了有效增强。综上,大坝变形组合预测模型具备出色的非线性信息挖掘与建模预测能力,可为大坝安全监测提供可靠依据。

基于Stacking特征增强多粒度联级Logistic的个人信用评估

主要针对广受关注的P2P网贷信用评估问题,利用机器学习方法提高申请人网贷违约预测准确率,研究出基于Stacking特征增强多粒度联级Logistic方法及其应用.所提分类器是一种混合模型,结合了Stacking集成学习和联级Logistic学习的思想.首先,通过网格搜索技术分别建立XGBoost,Catboost,LightGBM,AdaBoost以及Gradient Boosting模型,并筛选出适合的基评估器作为Stacking集成的初级学习器,logistic模型作为次级学习器,构建基于Stacking的多粒度扫描器,生成预测结果作为元特征,拼接成新特征数据.其次,通过新特征数据以及元特征在每级Logistic上的特征增强建立联级Logistic Regression模型,并且与现有的单一集成学习器和各基评估器在3个不同的P2P网贷信用评估数据集上进行对比.实验结果表明,通过AUC、准确率等指标对其进行评价,相比于各基评估器以及其他单一集成分类器,基于Stacking增强多粒度联级Logistic模型有较高的准确率,预测效果更优.

基于Stacking-Bert集成学习的中文短文本分类算法

由于word2vec、Glove等静态词向量表示方法存在无法完整表示文本语义等问题,且当前主流神经网络模型在做文本分类问题时,其预测效果往往依赖于具体问题,场景适应性差,泛化能力弱。针对上述问题,提出一种多基模型框架(Stacking-Bert)的中文短文本分类方法。模型采用BERT预训练语言模型进行文本字向量表示,输出文本的深度特征信息向量,并利用TextCNN、DPCNN、TextRNN、TextRCNN等神经网络模型构建异质多基分类器,通过Stacking集成学习获取文本向量的不同特征信息表达,以提高模型的泛化能力,最后利用支持向量机(support vector machine,SVM)作为元分类器模型进行训练和预测。与word2vec-CNN、word2vec-BiLSTM、BERT-TexCNN、BERT-DPCNN、BERT-RNN、BERT-RCNN等文本分类算法在网络公开的三个中文数据集上进行对比实验,结果表明,Stacking-Bert集成学习模型的准确率、精确率、召回率和F_(1)均为最高,能有效提升中文短文本的分类性能。