Generative adversarial network fairness analysis method, system and device and storage medium

Abstract

The invention belongs to the field of machine learning, and discloses a generative adversarial network fairness analysis method, system and device and a storage medium, and the method comprises the following steps: analyzing a to-be-analyzed attribute, and constructing an attribute classifier; analyzing whether the to-be-analyzed generative adversarial network can obtain a large number of generation samples; if yes, a large number of generation samples of the to-be-analyzed generative adversarial network are acquired and collected; otherwise, constructing a shadow model locally according to the characteristics of the to-be-analyzed generative adversarial network, optimizing the input noise of the to-be-analyzed generative adversarial network by using the shadow model, and obtaining and collecting a small number of generation samples of the to-be-analyzed generative adversarial network by using the optimized noise; and realizing fairness analysis of the to-be-analyzed generative adversarial network by using the attribute classifier and the attribute extraction function. According to the analysis method, two analysis processes based on a large number of samples and a small number of optimized samples are constructed according to the actual use scene of the to-be-analyzed generative adversarial network, and the applicability and success rate of fairness analysis of the to-be-analyzed generative adversarial network are effectively improved.

Description

technical field [0001] The invention belongs to the field of machine learning, and relates to a method, system, device and storage medium for fairness analysis of a generative confrontation network. Background technique [0002] In recent years, great breakthroughs have been made in related theoretical research on deep generative models. Among them, deep generative models represented by generative adversarial networks and variational autoencoders not only have powerful data distribution learning capabilities, but also provide the possibility for the generation of high-quality data samples. In addition to being a data augmentation technique for small-sample learning, deep generative models have also shown excellent performance in multiple types of information synthesis tasks such as video face swapping, music synthesis, and style transfer. While deep generative models have been successfully applied in key technical fields such as medical image reconstruction, portrait enhanc...

AI Technical Summary

Problems solved by technology

The current research work shows that even if it runs as a black box service, the training data or internal model information of the machine learning system may still be attacked in various forms and lead to privacy leakage. Typical examples include data privacy-oriented model reverse attacks, and Model Extraction Attacks for Model Privacy

Leakage of training data may directly violate the data privacy of users, and leakage of model information may cause economic losses to service providers. Attackers can even use this to analyze the weaknesses of the stolen model and launch further attacks, such as Anti-sample attack

[0004] Therefore, the current GAN strengthens the privacy protection of the training data set, which to a certain extent makes it impossible for users to determine whether their training data set is discriminatory in certain attributes, such as the data set in a certain attribute. equality, which in turn leads to the inability to effectively judge the fairness of the generative confrontation network, and the training process of some machine learning models depends on the samples generated by the generative confrontation network. The machine learning model trained by the data generated by the unfair generative confrontation network is likely to inherit The bias of sensitive attributes, such as age, gender, skin color and region, leads to serious discriminatory effects and serious fairness problems in machine learning models, which in turn raises concerns about its application in the real world

Images