A Spam Transaction attack is a kind of hostile attack activity specifically targeted against a Cryptocurrency Network. Traditional network intrusion detection methods lack the capability of automatic feature extraction for spam transaction attacks, and thus the detection efficiency is low. Worse still, these kinds of attack methods and the key intrusion behaviour process are usually concealed and submerged into a large number of normal data packages; therefore, the captured threat test samples are too small, which easily leads to insufficient training of detection model, low detection accuracy rate, and high false alarm rate. In this paper, a spam transaction intrusion detection model based on GRU(Gated Recurrent Unit) is proposed, which takes advantage of the excellent features of deep learning and uses repeated and multilevel learning to perform automatic feature extraction for network intrusion behaviour. The model has extremely high learning ability and massive data processing ability. Moreover, it has a quicker and more accurate spam transaction attack detection ability than traditional intrusion detection algorithms. Additionally, a generation method of spam transaction-samples based on WGAN-div is proposed, which obtains new samples by learning training samples and solves the problems of insufficient original samples and unbalanced samples. A series of experiments were performed to verify the proposed models. The proposed models can distinguish between normal and abnormal transaction behaviours with an accuracy reaching to 99.86%. The experimental results indicate that the proposed models in this paper have higher efficiency and accuracy in detecting spam transaction attacks, which provides a novel and better idea for research of spam transaction attack detection systems.

垃圾邮件交易攻击是一种专门针对加密货币网络的敌对攻击活动。传统的网络入侵检测方法缺乏针对垃圾邮件交易攻击的自动特征提取能力，检测效率低。更糟糕的是，这些攻击方法和关键入侵行为过程通常被隐藏并淹没在大量正常数据包中。因此，捕获的威胁测试样本太小，容易导致检测模型训练不足，检测准确率低，误报率高。本文提出了一种基于GRU（门控循环单元）的垃圾邮件交易入侵检测模型，它利用深度学习的出色功能，并使用重复和多层学习来针对网络入侵行为执行自动特征提取。该模型具有极高的学习能力和海量数据处理能力。而且，与传统的入侵检测算法相比，它具有更快，更准确的垃圾邮件交易攻击检测能力。另外，提出了一种基于WGAN-div的垃圾邮件交易样本生成方法，该方法通过学习训练样本获得新样本，解决了原始样本不足和样本不平衡的问题。进行了一系列实验以验证提出的模型。所提出的模型可以区分正常交易行为和异常交易行为，准确率高达99.86％。