【作者】 王喆；

【导师】 郭东伟；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2018， 硕士

【摘要】 21世纪是互联网的时代,科学技术正在蓬勃发展,促进了人类社会飞速前进。而在互联网时代的众多研究方向中,人工智能脱颖而出,成为了近年最火热的研究项目。各个互联网公司,研究实验室,学校都在积极展开对人工智能的研究工作。其中在图像识别领域,卷积神经网络凭借着其高准确率成为受捧的网络结构。但是Nguyen在2015年发现了一个事实:深度神经网络很容易受到欺骗,会用极高的置信度去分类某一个不可识别的图片。深度神经网络虽然在人脸识别和数字识别中有着杰出的准确率,但是科研人员可以很容易的利用梯度上升法产生一些人眼完全无法识别的图像,但是最先进的深度神经网络模型却以99.99%的置信度认为这些图像是可以识别的,并且给出其“分类”。特别的,我们在一个已经用MNIST数据库或者ImageNet数据库完全训练好,并且表现很好的卷积神经网络中,用遗传算法或者梯度上升法生成使得输入图像能分类到对应标签且置信度最大的输入图像,会发现这些图像人眼并不可以辨认。本文首先介绍了双向联想记忆神经网络(BAM),卷积神经网络(CNN),t-SNE的原理和作用,然后提出了一个基于卷积神经网络和t-SNE的双向联想记忆神经网络模型。在该模型中,将CNN和t-SNE提取到的特征向量送入BAM进行联想记忆,当从BAM的一端输入一种特征向量时,另一端会输入联想到的结果,并且具有一定容错性。利用两种特征互相牵制,将联想到的结果特征向量和正确结果特征向量进行相似度判断,并设置相似度可变阈值n。然后采用MNIST数据库和100个用梯度上升法得到的欺骗数据对模型进行了测试实验。试验结果表明,在没有欺骗数据的时候,CNN的准确率非常理想,达到了99.8%,但是用欺骗数据测试,所有数据均被以超过98%的置信度划分为某一个标签;在我们的模型中,阈值设置为0.25的时候,这部分数据只有8%依然被划分到对应标签,其他数据均被识别为欺骗数据。可见我们的模型在对抗欺诈数据上有着明显的提升,能更好的保证识别鲁棒性。更多还原

【Abstract】 The 21 st century is the era of the Internet.Science and technology are booming,promoting the rapid progress of human society.In the Internet era of many research directions,artificial intelligence come to the fore,has become the hottest research project in recent years.All Internet companies,research laboratories,schools are actively engaged in research on artificial intelligence.Among them,in the field of image recognition,the convolutional neural network has become a favored network structure with its high accuracy.But Nguyen found in 2015 a fact that deep neural networks are easily tricked into classifying unidentifiable pictures with high confidence.Although deep neural networks have outstanding accuracy in face recognition and digital recognition,researchers can easily make use of gradient ascent to produce images that are completely unrecognizable by human eyes.However,the most advanced deep neural network models use 99.99 The confidence of% considers these images to be recognizable and gives their "classification." Specifically,in a convolutional neural network that has been fully trained and performed well with the MNIST database or ImageNet database,genetic algorithms or gradient ascending methods are used to generate inputs that allow the input images to be sorted to the corresponding labels with the highest confidence The images,you will find these images are not recognized by the human eye.In this paper,the principle and function of bidirectional associative memory neural network(BAM),convolutional neural network(CNN)and t-SNE are introduced firstly.Then a bidirectional associative memory neural network model based on convolutional neural network and t-SNE is proposed.In this model,the feature vectors extracted by CNN and t-SNE are sent to BAM for associative memory.When an eigenvector is input from one end of BAM,the other end will input the result of association and have some fault tolerance.The two features are used to check each other,and the similarity between the associative result eigenvector and the correct result eigenvector is judged,and the similarity similarity threshold n is set.Then the MNIST database and 100 deception data obtained by the gradient ascent method were used to test the model.The test results show that the accuracy of the CNN is very good at 99.8% without cheating the data,but with the cheating data test,all the data are classified as a single label with more than 98% confidence.In our model,the threshold is set to 0.25,only 8% of this part of the data is still divided into the corresponding label,the other data are identified as deceptive data.It can be seen that our model has obvious improvement in anti-fraud data and can better ensure the recognition robustness.更多还原