Abstract: Financial institutions such as banks need to verify the authenticity of users when verifying their loan needs。 And a common method to confirm the user's identity is investigating several users' personal information by telephone。 Impostors can obtain user information through illegal channels because of increasing information security issues and impersonate real users to defraud the loan。 Everyone's speaker features are unique and unforgeable， so it’s feasible using voice data to authenticate。 Therefore， we introduce a loan fraud algorithm in this article。 We use users' telephone recordings to extract their voiceprint and build a speaker model library。 And so we found that users collections with similar voiceprint in order to identify fraud behavior。 At the end we comparing performance of GMM-UBM and I-vector based loan fraud algorithm。

Key words: loan fraud algorithm, speaker recognition, GMM-UBM, I-vector

摘要： 银行等金融机构在用户贷款时需要核实用户身份的真实性，常见的方法是通过电话问询用户个人信息的方式来确认身份。日益严重的信息安全问题导致骗贷人可以通过非法途径获取用户信息，冒充真实用户来骗取贷款。本文介绍了一种防骗贷算法，从用户的电话录音中提取用户的声纹特征，建立用户的声纹模型库，发现具有高相似度的声纹模型集，鉴别出冒充不同用户身份的骗贷者。最后测试并比较了基于GMM-UBM和I-vector模型的最大团防骗贷算法性能。

关键词: 防骗贷算法, 说话人识别, GMM-UBM, I-vector