Abstract: With the wide application of deep learning technology in image recognition,natural language processing and financial predicting,once there is bias in analysis results,it will cause negative impacts both on individuals and groups,thus any effects on its performance it is vital to enhance the fairness of the model without affecting the perfomance of deep learning model.Biased information about data is not only sensitive attributes,and non-sensitive attributes will also contain bias due to the correlation among attributes,therefore,the bias cannot be eliminated when debiasing algorithms only consider sensitive attributes.In order to eliminate the bias in the classification results of the deep learning model caused by the correlated sensitive attributions in the data,this paper proposes a data debiasing method based on the generative adversarial network.The loss function of the model combines the fairness constraints and the accuracy loss,and the model utilizes adversarial code to eliminate bias to generate debiased dataset,then with the alternating gaming training of the generator and the discriminator to reduce the loss of the no-bias information in the dataset,and the classification accuracy is ensured while the bias in the data is eliminated to improve the fairness of the subsequent classification tasks.Finally,data debiasing experiments are carried out on several real-world dataset to verify the effectiveness of the proposed algorithm.The results show that the proposed method can effectively decrease the bias information in datasets and generate datasets with less bias.

Key words: Adversarial training, Data debiasing, Deep learning, Generative adversarial networks