PhD Proposal: On the Bright and Dark Sides of Generative Adversarial Networks

Talk
Ning Yu
Time: 
12.10.2020 15:00 to 17:00
Location: 

Remote

Photorealistic image generation has increasingly become reality, benefiting from the invention of generative adversarial networks (GANs) and its successive breakthroughs. Both the bright side and the dark side of this technique have attracted substantial attention. On the bright side, GANs have been popularized into extensive computer vision applications. In the first part of this presentation, we demonstrate its successful application in controllable texture interpolation. On the dark side, there rise strong concerns on how GANs can be misused to spoof sensors, generate deep fakes, and enable misinformation at scale. We consequently present in the second part our study of learning GAN fingerprints towards deep fake detection and attribution. Further concerns come from the fact that GANs exacerbate the imbalance in training data and generate biases against minority groups. Therefore, in the third part, we formalize the problem of minority inclusion as one of data coverage, and show improved coverage by harmonizing adversarial training with reconstructive generation.The next stage of my Ph.D. research sticks to the bright and dark sides of GANs. On one hand, we propose to mitigate two important limitations of the state-of-the-art GAN technique: (a) local and non-adaptive spatial dependencies in image generation, and (b) overfitted feature representation in adversarial training. On the other hand, we propose to extend the learning of GAN fingerprints towards a proactive solution of deep fake attribution: embedding artificial fingerprints into GAN models so as to enable responsible disclosure when publicizing models.Examining Committee:

Chair: Dr. Larry Davis Dept rep: Dr. Matthias Zwicker Members: Dr. David Jacobs Dr. Abhinav Shrivastava