North Coast Synthesis Ltd.

Dog-whistle GANs

2025-05-21, access: Public | Next ▶

basics training security theory image GAN Going back here to the original paper on GANs, Goodfellow et al. from 2014. It's a short paper and a simple concept: you train your generative model in tandem with another model (probably quite a simple one) that tries to distinguish the main model's output, from training data. The "discriminator" model is basically being used as a glorified loss function. The paper's point is simply that that works.

In the talk I'm interested in pushing in the direction of watermarking: what does the GAN technique mean for "plagiarism detectors" and similar? Can we build a model specifically with the goal of having it be detectable? Can we play rock-paper-scissors-lizard-Spock with five models trying to fake each other out? And so on.

Access level to read comments: Public
Access level to write comments: Free account (logged in)