adversarial

Definition

Two distinct but related senses. In training, it means pitting two components against each other, as in a GAN where a generator learns against a discriminator. In security and robustness work, it means inputs crafted to make a model fail — adversarial examples, adversarial prompts, jailbreaks. What unites them is optimizing against a worst case rather than an average one.

Explainers using this term

• The Rise and Fall of GANs — An Invention Trained by Rivalry, and Why Diffusion WonGenerative Adversarial Networks