The discriminator of ContraGAN discriminates the authenticity of given samples and minimizes a contrastive objective to learn the relations between training images.

In this work, our goal is to fill this gap.

Our key idea is to construct a test-sample-specific auxiliary task to update the model before applying it to the sample. Specifically, we synthesize pseudo-training samples from each test image and create a test-time training objective to update the model.

Despite achieving remarkable performance, deep graph learning models, such as node classification and network embedding, suffer from harassment caused by small adversarial perturbations.