Leveraging Contaminated Datasets to Learn Clean-Data Distribution with Purified Generative Adversarial Networks

• DOI: 10.48550/arxiv.2302.01722
• Publication Date: 2023-01-01
• Authors: Bowen Tian; Qinliang Su; Jianxing Yu
• Keywords: Generative adversarial network

Generative adversarial networks (GANs) are known for their strong abilities on capturing the underlying distribution of training instances. Since seminal work GAN, many variants GAN have been proposed. However, existing GANs almost established assumption that dataset is clean. But in real-world applications, this may not hold, is, be contaminated by a proportion undesired When such datasets, will learn mixture desired and instances, rather than data only (target distribution). To target from two purified generative (PuriGAN) developed, which discriminators augmented with capability to distinguish between instances leveraging an extra solely composed contamination We prove under some mild conditions, proposed PuriGANs guaranteed converge Experimental results several datasets demonstrate able generate much better images comparable baselines when trained datasets. In addition, we also usefulness PuriGAN downstream applications applying it tasks semi-supervised anomaly detection PU-learning. show deliver best performance over both tasks.