Results of paper "Bias-Free Federated GAN"
We present the results of our paper "Bias-Free Federated GAN here. We demonstrate the capabilities of our model in the following datasets:
We show the results of both federated SpecGAN and Bias-Free Federated SpecGAN on both the datasets. Please use headphones for better quality. We also show results on MNIST, Fashion MNIST, Cifar 10 and FairFace Datasets.
We consider two classes "Zero" and "Eight" in SC09 dataset. Client 0 has datapoints of class "Zero" and Clients 1 and 2 have datapoints of class "Eight". Note that in the normal Federated version, the federated model is not able to generate datapoints of the minority class "Zero", while in the Bias-Free Federated version, the federated model can also synthesize minority classes.
We consider two classes "Down" and "Up" in Mini Speech Commands dataset. Client 0 has datapoints of class "Down" and Clients 1 and 2 have datapoints of class "Up". Note that in the normal Federated version, the federated model is not able to generate datapoints of the minority class "Down", while in the Bias-Free Federated version, the federated model can also synthesize minority classes.