基于特征优化生成对抗网络的在线交易反欺诈方法研究  

An Generative Adversarial Network of Online Transaction Anti-fraud Method Based on Feature Optimization

作　　者：张浩 康海燕[1] ZHANG Hao;KANG Haiyan(Department of Information Security, Beijing Information Science and Technology University, Beijing 100192,China)

机构地区：[1]北京信息科技大学信息安全系,北京100192

出　　处：《郑州大学学报（理学版）》2022年第1期69-74,87,共7页Journal of Zhengzhou University:Natural Science Edition

基　　金：教育部人文社科项目(20YJAZH046);研究生课程建设项目(2020YKJ17);国家自然科学基金项目(61370139)。

摘　　要：为了降低在线交易欺诈数据的不平衡性对欺诈检测效果的影响,提出了一种基于特征优化生成对抗网络的在线交易反欺诈方法。该方法建立了WGAN网络包括生成模型和判别模型,对数据进行Key特征选取,在数据生成过程中进行Gumbel-softmax技巧采样输出,优化生成数据质量和提高训练稳定性;交替训练判别模型和生成模型直至模型收敛;接着将收敛的生成模型作为样本生成器生成少数类样本对原始数据进行平衡处理;利用平衡处理后的数据训练分类模型并进行模型评估。通过实验证明,该方法生成数据的效果优于SMOTE及其变种方法。In order to reduce the impact of the imbalance of online transaction fraud data on fraud detection,an online transaction anti-fraud method based on feature optimization generation counter network was proposed.The WGAN network was established,including the generation model and the discrimination model.Key features were selected for the data,and Gumbel softmax was used to sample and output the data generated by the model to optimize the quality and the stability.The discriminant model and generating model were trained alternately until the model converged.The convergent generation model was used as a sample generator to generate a few class samples to balance the original data.The classification model was trained and evaluated by using the balanced data.Experimental results showed that this method was better than SMOTE and its variants in data generation.

关 键 词：交易反欺诈 生成对抗网络(GAN) Wasserstein GAN(WGAN) Gumbel-softmax 不平衡数据 

分 类 号：F830[经济管理—金融学] TP183[自动化与计算机技术—控制理论与控制工程]