信贷自动审批模型的对抗攻击风险研究

Research on adversarial attack risk of automatic credit approval model

近期,银行等金融机构引进自动信贷审批系统来取代传统的人工审批,而自动信贷审批系统在何种程度上会受到对抗样本的攻击有待研究。通过实验对信贷对抗样本攻击的问题进行了验证。首先,基于申请人的信贷数据对XGBoost模型进行训练,预测申请人行为,并选择原始样本。其次,使用“非违约申请人”对改进的GAN模型进行训练,并用于生成特征值,通过修改原始样本以构建对抗样本,使得修改后的特征值接近于“非违约申请人”密集分布的特征值。最后,使用训练好的XGBoost模型将对抗样本进行分类。在实验中生成的对抗样本可以混淆XGBoost模型。当修改后的特征值的数量增加时,对抗样本的生成率总体呈上升趋势。实验验证,对抗样本的攻击将对自动信贷审批系统造成安全风险。

Recently,many banks and other financial institutions have introduced automated credit approval systems to replace the traditional manual approval.Automatic credit approval system is vulnerable to attack from adversarial examples.This paper presents an experiment of credit adversarial examples attack.Firstly,the XGBoost model is trained based on all applicants'credit data to predict applicant behavior and select the original sample.Secondly,the′Non-default applicant'is used to train the improved GAN model and generate eigenvalues,and then some features of the original sample are modified to construct the confrontation sample.The modified eigenvalue is close to the eigenvalue of the dense distribution of‘Non-default applicant'.Finally,the trained XGBoost model is used to classify the confrontation examples.We find that in the experiment,the generated confrontation examples can confuse the XGBoost model.When the number of modified eigenvalues increases,the generation rate of confrontation examples is on the rise.In short,the experiment shows that the attack from adversarial examples will pose security risks to the automatic credit approval system.