基于歧视对搜索的神经网络公平性测试  

Fairness testing of neural network based on discrimination in search

作　　者：张煌 叶阿勇[1] ZHANG Huang;YE Ayong(Fujian Normal University,Fuzhou 350000,China)

机构地区：[1]福建师范大学,福州350000

出　　处：《计算机应用文摘》2025年第1期163-165,共3页

基　　金：移动网络中面向位置信息生命周期的隐私保护关键技术(61972096)。

摘　　要：近年来,神经网络模型在金融、医疗和司法等关键决策领域的应用日益广泛。然而,这些模型可能因数据中的偏差而导致不公平决策,尤其是在涉及种族、性别等敏感特征时。为了解决这一问题,文章提出了一种基于歧视对搜索的公平性测试算法,通过混合整数线性规划形式化地检测神经网络模型的个体公平性。实验结果表明,该方法能够在保证模型性能的同时,有效评估模型的公平性。In recent years,the application of neural network models in key decision-making fields such as finance,healthcare,and justice has become increasingly widespread.However,these models may lead to unfair decisions due to biases in the data,especially when sensitive features such as race and gender are involved.To address this issue,the article proposes a fairness testing algorithm based on discrimination based search,which formally detects the individual fairness of neural network models through mixed integer linear programming.The experimental results show that this method can effectively evaluate the fairness of the model while ensuring its performance.

关 键 词：神经网络模型 个体公平 歧视对搜索 公平性测试 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]