Method, system, device and medium for testing fairness of deep neural network model

Abstract

The invention belongs to the field of machine learning, and discloses a deep neural network model fairness testing method, system, equipment and medium, including obtaining several samples, querying the neural network model to be tested to obtain the prediction results of each sample, and using known prediction results The samples are clustered into sample clusters, and the surrogate models are trained separately, based on the corresponding surrogate models on each sample cluster, a new set of seed samples is generated, and disturbances are applied to the extracted seeds, so that they can be obtained with a high probability It becomes a sample that violates the fairness condition, and based on multiple perturbations to the currently discovered sample that violates the fairness condition, more samples that violate the fairness condition are found to fully verify the fairness of the deep neural network model. This method has few restrictions, which fully meets the practical limitations of the actual application that the model cannot be obtained, the number of visits is limited, and only the output prediction results can be obtained. The overall performance is efficient, and the task of black-box fairness testing is well realized.

Description

technical field [0001] The invention belongs to the field of machine learning, and relates to a deep neural network model fairness testing method, system, equipment and medium. Background technique [0002] At present, deep neural network (DNN) is gradually being widely used in various fields of life, such as face recognition, automatic driving, medical diagnosis and recommendation system, etc., and has shown amazing performance and potential. Still, its reliability and security have flaws. In some application backgrounds with social impact, such as credit evaluation, crime prediction, etc., an ideal attribute of DNN is fairness, that is, there is no discriminatory prediction behavior. In practical applications, DNN violates fairness, ranging from individual Users have a real impact, and at worst, it will make the whole society deviate further from the public's expectations of fairness. Fairness detection for DNNs has become a research direction that has gradually attracte...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to overcome the fact that the existing DNN fairness testing method in the above-mentioned prior art cannot well satisfy the fairness under the realistic constraints of limited access to models, including limited number of model visits and limited query information. In view of the shortcomings of the fairness test, a deep neural network model fairness test method, system, equipment and medium are provided

Images