【作者】 曹雷；

【导师】 曹汉强；

【作者基本信息】 华中科技大学 ， 信息与通信工程， 2017， 硕士

【摘要】 近年来,深度神经网络在语音识别和图像分类方面取得极大的成就,在全世界掀起了一股深度学习的热潮。信息安全关乎都我们个人的核心利益与隐私,对我们每个人都至关重要。目前深度学习方面在信息安全领域还涉及得比较少,主要还是应用在模式分类上。本文从一个新的方向,探索了深度学习在信息安全领域的研究。根据深度神经网络(DNNs)与人类视觉在图像分类上的区别,本文提出了一种新的加密解密算法,本文的工作具有重要的理论意义与实用价值。本文首先详细分析了算法的原理:经过进化算法(EAs)梯度上升处理后的图片,人眼对其完全没法识别,但是经过训练的特定的深度神经网络能够以高于99%的置信度将其分类,经过处理的图片就是很好的信息隐藏方式,分类的结果就是我们需要加密或者隐藏的信息。同时对算法的可行性与适用范围进行了分析。然后利用Caffe搭建并训练了深度神经网络模型,得到并提取了网络的权值作为算法的密钥。利用Sferes进化算法框架来生成需要的噪声图片,这些图片通过DNNs分类得到的结果就是解密过程,可以得到我们需要的明文。实验结果表明,该算法在一些信息(比如数字和字母)的加密解密可以做到很高的安全性与解密正确率。同时,由于分类标签是我们需要加密的明文,我们可以把分类标签设置成我们需要加密的明文,就可以实现对不同种类的信息的加密隐藏。同时,本文对算法的性能与安全性进行了详细地分析。更多还原

【Abstract】 Deep neural networks(DNNs)have recently been achieving state-of the-art performance on a variety of speech recognition and image classification problems,and set off a boom of deep learning all over the world.Lots of scientific research fields have fallen into Deep Learning to solve their problems.At the same time,we are living in information age,and must deal with a huge number of information such as texts,images,voices and videos every day,which are the key to our benefits and privacy.At present,Deep Learning refers less to the field of information security,and apply to pattern-recognition tasks.This paper tries to explore the application of the Deep Learning on the information security from a different direction.This paper proposes a new encryption-decryption algorithm on the basis of the difference between CNNs and human vision on the image classification.The work in this paper possesses important theoretical significance and practical value.Firstly,this paper analyzes detailedly the theory: an image,which has been disposed by Evolutionary Algorithms or gradient ascent,is imperceptible to human completely.But a trained DNNs believe to be recognizable objects with 99% confidence.This disposed image is perfect way to encrypt or hide information.,the feasibility and the application of the algorithm.Then,we set up and train the DNNs model by the Caffe,and extract the weights of this model as the secret key of algorithm.We also produce images we need by Sferes2,which is a high-performance,multi-core,lightweight,generic C++ framework for Evolutionary Algorithms,and the process of decryption of the algorithm is that CNNs classifies these images and gives the classification results.The experiments show that this algorithm can be very security on encryption and very accurate on decryption to some information such as digits and letters.At the same time,on account of that the classification labels is the information need to encrypt,we can set the information as the classification labels and encrypt different information.At the same time,this paper analyzes the performance and security of this algorithm detailedly.更多还原