Abstract: A novel fraud call recognition model named GLPMAM was proposed to address the problem that the correlation between users is rarely considered in traditional fraud call recognition models. The call-making relationships between phone users were modeled as a graph to consider the correlation between them, and user features were extracted by combining multi-layer perceptron (MLP) and convolutional neural network (CNN). Finally, the multi-channel attention mechanism was applied to further improve the accuracy. Experiments on public datasets provided by the first Digital Sichuan Innovation Competition showed that the accuracy of GLPMAM reached over 99%, which was 7% higher than the existing fraud call recognition model. The ablation experiment verified the important role of graph construction in improving accuracy. Experimental results demonstrated that GLPMAM could effectively improve the accuracy of fraud call recognition in practical applications, and the model running speed could adequately meet the needs of practical applications.

Key words: fraud phone identification, graph, multi-layer perceptrons, convolutional neural network, multi channel attention mechanism

摘要： 针对传统诈骗电话识别模型中较少考虑用户间的相关性的问题，提出了一种新颖的诈骗电话识别模型GLPMAM，该模型通过将电话用户之间的接打电话关系建模为图来考虑电话用户之间的相关性，并结合多层感知机（MLP）和卷积神经网络（CNN）提取用户特征，最后通过多通道注意力机制进一步提升准确率。在公开数据集首届数字四川创新大赛提供的公开数据集上的实验表明，GLPMAM的准确率达到99%以上，较现有的诈骗电话识别模型提升了7%以上，并通过消融实验验证了图的建构对于准确率提升的重要作用。实验结果表明，GLPMAM在实际应用中能够有效提升诈骗电话识别准确率，且模型运行速度完全能够满足实际应用需求。

关键词: 诈骗电话识别, 图, 多层感知机, 卷积神经网络, 多通道注意力机制