페어 간 하고 XAI

Research project at the Graduate School of Data Science at Seoul National University. The overarching aim of the project was to attempt to neutralize dataset biases by generating synthetic datasets with Generative Adversarial Networks (GAN) and reducing the performance gap on classification models trained on these datasets. The project used a so called FairGAN (Xu et al., 2018), which is suggested to produce less biased and more diverse data. The goals of the project was to evaluate the fairness and bias of data generated by a FairGAN compared to data generated by a default GAN and to evaluate performance in terms of fairness and bias of a face recognition classifier trained with a FairGAN data compared to trained with default GAN data.

In this project, we synthesized 2 different datasets of human-faces, one generated by a regular GAN and one by a FairGAN by training on UTKface dataset. We then compared the diversity distribution and statistics of 3 datasets (2 synthesized datasets & original training set) with e.g. TSNE- and CDF- analysis and trained a gender classification model on the 2 synthesized datasets to compare performance. Finally, we used explainability methods like LIME, SHAP and GradCAM to visually analyze and explain the behavior of the classifier with respect to each dataset. We evaluated fairness based on explainability visualisations as well as qualitative analysis (Sample explanations, REVISE).