【摘要】 由于IPv6拥有庞大的地址空间，基于现有网络速度和硬件计算能力，难以实现全球IPv6地址扫描。通过地址生成算法来预测网络中可能出现的IPv6地址，随后将预测地址作为扫描的目标，可以达到IPv6地址快速扫描的目的。文中通过分析IPv6地址结构和分配方式来探索潜在的分配模式，结合已有的传统语言模型和目标生成算法，提出了一种基于深度学习的算法6LMNS,来预测潜在的活跃IPV6地址。6LMNS首先通过地址向量空间映射模型Add2vec来构建具有一定语义关系的IPv6地址词向量空间；随后基于Transformer构建语言训练模型GPT-IPv6,以此来估计IPv6地址词向量序列的概率分布；最后引入核心采样替代传统贪心搜索解码，完成活跃地址的生成。经验证，与其他语言模型和目标生成算法相比，6LMNS生成的地址拥有更好的多样性以及更高的活跃率。更多还原

【Abstract】 The huge address space of IPv6 makes it difficult to achieve a global IPv6 address scan based on the existing network speed and hardware computing power.Fast IPv6 address scanning can be achieved by using address generation algorithms to predict the possible IPv6 addresses in the network and subsequently using the predicted addresses as the targets of scanning.This paper explores potential allocation patterns by analyzing IPv6 address structures and allocation methods, and proposes a deep learning-based algorithm 6LMNS to predict potentially active IPV6 addresses by combining existing traditional language models and target generation algorithms.6LMNS first constructs IPv6 address word vector spaces with certain semantic relationships through the address vector space mapping model Add2vec.Subsequently, the language training model GPT-IPv6 is constructed based on Transformers to estimate the probability distribution of IPv6 address word sequences.Finally, nucleus sampling is introduced instead of traditional greedy search decoding to complete the generation of active addresses.It is verified that the addresses generated by 6LMNS have better diversity as well as higher activity rate compared with other language models and target generation algorithms.更多还原