基于多特征自适应融合的区块链异常交易检测方法

Block-chain abnormal transaction detection method based on adaptive multi-feature fusion

针对智能检测模型的性能受限于原始数据(特征)表达能力的问题,设计了一种残差网络结构ResNet-32用于挖掘区块链交易特征间隐含的关联关系,自动学习包含丰富语义信息的高层抽象特征。虽然浅层特征区分能力弱,但更忠于原始交易细节的描述,如何充分利用两者的优势是提升异常交易检测性能的关键,因此提出了特征融合方法自适应地桥接高层抽象特征与原始特征之间的鸿沟,自动去除其噪声和冗余信息,并挖掘两者的交叉特征信息获得最具区分力的特征。最后,结合以上方法提出区块链异常交易检测模型(BATDet),并通过Elliptic数据集验证了所提模型在区块链异常交易检测领域的有效性。

Aiming at the problem that the performance of intelligent detection models was limited by the representation ability of original data(features), a residual network structure ResNet-32 was designed to automatically mine the intricate association relationship between original features, so as to actively learn the high-level abstract features with rich semantic information. Low-level features were more transaction content descriptive, although their distinguishing ability was weaker than that of the high-level features. How to integrate them together to obtain complementary advantages was the key to improve the detection performance. Therefore, multi feature fusion methods were proposed to bridge the gap between the two kinds of features. Moreover, these fusion methods can automatically remove the noise and redundant information from the integrated features and further absorb the cross information, to acquire the most distinctive features. Finally, block-chain abnormal transaction detection model(BATDet) was proposed based on the above presented methods, and its effectiveness in the abnormal transaction detection is verified.