【作者】 康明；

【导师】 陈文； 杨韬；

【作者基本信息】 四川大学 ， 网络与信息安全（专业学位）， 2023， 硕士

【摘要】 在物联网(Internet of Things,IoT)时代,安全是决定物联网系统生存能力的关键因素。而传统的加解密码算法因计算过程复杂,难以适用于资源受限的物联网环境,如RFID系统、WSN传感网络等。因此,人们为IoT环境设计了一系列低功耗、低资源占用,且满足一定加密强度要求的轻量级加密算法。如美国国家安全局(National Security Agency,NSA)设计了用于物联网加密的轻量级分组加密算法SIMON,旨在满足低功耗和有限栅门器件的硬件要求下,实现通信数据的保密安全传输。随着5G、物联网等技术的迅速普及,追求高效率和低功耗的SIMON等物联网加密算法将得到大量应用,同时也势必面临严峻的安全考验。因此,对各类轻量级的加密算法进行安全性评估,以提前发现算法的安全性漏洞便尤为重要。而差分分析、线性分析等传统的密码分析方法都需要差分、线性掩码等先验知识和经验推导。近年来,深度学习被广泛应用于科学技术领域进行数据分析和知识推理。2019年,Gohr等首先提出用神经网络替代传统的差分分布表,并成功应用于分组加密算法SPECK的安全性评估过程。随后,研究者基于Gohr的模型,开始从差分选择、神经网络网络结构优化等方面改进研究,并应用于不同分组密码算法的安全评估中。目前,神经网络区分器在较高轮次的迭代加密算法的安全性评估中还难以适用,存在密文随机性随加密轮次增加、高概率的差分和线性掩码特征等关键参数搜索困难等问题;此外,在密钥恢复过程中,存在有效的密文样本搜索代价高,密钥恢复效率较低等问题。针对上述问题,本文以SIMON为例,进行基于深度学习的SIMON差分-线性分析,将深度学习与传统的差分-线性密码安全评估相结合,探索自动化的轻量级分组密码安全评估方法,主要开展了以下工作:(1)本文将神经网络与差分-线性密码分析方法相结合,提出密文的差分-线性特征表示方法,结合密文的差分特征和线性掩码特征筛选密文的高概率特征路径,提高了神经网络区分器的密文区分能力。与Gohr的神经网络差分区分器的对比实验结果表明,神经网络差分-线性区分器在SIMON32/64相同8轮加密轮数下,区分精度提升了28.74%;(2)提出密文敏感度分析方法,以识别对神经网络差分-线性区分器性能有显著影响的敏感密文比特位;基于识别出的敏感位序列,利用VNS变邻域搜索算法,实现了差分-线性掩码的自动化搜索;(3)提出代表性样本集生成策略,生成代表性密文样本用于密钥恢复攻击,提升了神经网络区分器在密钥恢复攻击的效率,同时降低了密钥恢复攻击的空间复杂度。更多还原

【Abstract】 In the Internet of Things(Internet of Things,IoT)era,security is the key factor to determine the viability of IoT systems.The traditional encryption and decryption code algorithms are difficult to be applied to the resource-constrained IoT environment,such as RFID systems and WSN sensor networks,due to the complex computation process.Therefore,a series of lightweight encryption algorithms have been designed for IoT environments with low power consumption,low resource consumption,and meeting certain encryption strength requirements.For example,the U.S.National Security Agency(National Security Agency,NSA)designed SIMON,a lightweight packet encryption algorithm for IoT encryption,which aims to achieve confidential and secure transmission of communication data under the hardware requirements of low power consumption and limited gate devices.With the rapid spread of 5G,IoT and other technologies,IoT encryption algorithms such as SIMON,which pursues high efficiency and low power consumption,will be used in large numbers,and will also inevitably face severe security tests.Therefore,it is especially important to evaluate the security of various lightweight encryption algorithms in order to discover the security vulnerabilities of the algorithms in advance.Traditional cryptanalysis methods such as differential analysis and linear analysis require a priori knowledge and empirical derivation such as differential and linear masks.In recent years,deep learning has been widely used in science and technology for data analysis and knowledge inference.2019,Gohr et al.first proposed to replace the traditional differential distribution table with a neural network and successfully applied it to the security evaluation process of the packet encryption algorithm SPECK.Subsequently,based on Gohr’s model,researchers started to improve their research in terms of differential selection and optimization of neural network network structure and applied it to the security evaluation of different grouped cryptographic algorithms.At present,the neural network distinguisher is still difficult to be applied in the security evaluation of higher rounds of iterative encryption algorithms,and there are problems such as the increase of ciphertext randomness with the number of encryption rounds,the difficulty of searching key parameters such as differential and linear mask features with high probability;in addition,there are problems such as the high cost of searching effective ciphertext samples and the low efficiency of key recovery in the key recovery process.To address the above problems,this thesis conducts a deep learning-based differential-linear analysis of SIMON as an example,combines deep learning with traditional differential-linear cryptographic security evaluation,and explores an automated lightweight packet cryptographic security evaluation method,mainly carrying out the following work:(1)In this thesis,we combine neural networks with differential-linear cryptanalysis methods,propose differential-linear feature representation of ciphertexts,combine differential features of ciphertexts and linear mask features to screen high probability feature paths of ciphertexts,and improve the ciphertext differentiation capability of neural network differentiators.The comparison experimental results with Gohr’s neural network differential distinguisher show that the neural network differential-linear distinguisher improves the distinguishing accuracy by 28.74% under the same 8 encryption rounds of SIMON32/64.(2)Proposed a ciphertext sensitivity analysis method to identify sensitive ciphertext bits that have a significant impact on the performance of the neural network differential-linear distinguisher;based on the identified sequence of sensitive bits,the VNS variable neighborhood search algorithm is used to automate the search of differential-linear masks.(3)A representative sample set generation strategy is proposed to generate representative ciphertext samples for key recovery attacks,which improves the efficiency of the neural network distinguisher in key recovery attacks and reduces the space complexity of key recovery attacks at the same time.更多还原