基于仿生模式识别的用户概貌攻击集成检测方法

Ensemble Approach for Detecting User Profile Attacks Based on Bionic Pattern Recognition

针对有监督方法在检测用户概貌攻击时准确率不高的问题,通过引入仿生模式识别理论和集成学习技术提出一种集成检测方法.首先,通过计算被覆盖直线段与最近邻真实概貌的距离,提出一种自适应神经元超球半径计算算法,为每个神经元确定合适的超球半径;然后利用该超球半径对现有的一个3层神经网络进行重新设计,使其能够对攻击概貌样本进行更合理覆盖,以提高分类性能;最后,提出一种用户概貌攻击集成检测框架,通过组合多种攻击类型,利用提出的基训练集生成算法建立不同的基训练集,以训练新设计的神经网络生成基分类器,基于信息论得分(information theoretic score,ITS)算法提出一种选择性集成检测算法对基分类器进行筛选,并采用多数投票策略融合基分类器的输出结果.在MovieLens和Netflix两个不同规模的真实数据集上的实验结果表明,所提出的集成检测方法能够在保持较高召回率的条件下有效提高用户概貌攻击检测的准确率.

The supervised approaches suffer from low precision when detecting user profile attacks.Aiming at this problem,an ensemble detection approach is proposed by introducing bionic pattern recognition theory and ensemble learning technology.Firstly,through calculating the distance between a covered line segment and its nearest genuine profile,an adaptive calculational algorithm is proposed to adaptively assign a proper radius of hypersphere for each neuron.Secondly,the assigned radius is used to redesign an existing constructive neural network to make it more reasonable to cover attack profiles so as to improve its classification capability.Finally,an ensemble framework is proposed to detect user profile attacks.To create diverse base training sets,a base training set generation algorithm is proposed by combining various attack types.These base training sets are used to train the redesigned neural network in order to generate base classifiers.Based on ITS algorithm,a selective ensemble detection algorithm is proposed to select parts of base classifiers and the majority voting strategy is used to integrate the outputs of these base classifiers.The experimental results on two different scale of real datasets,MovieLens and Netflix,show that the proposed approach can effectively improve the precision under the condition of holding a high recall.