Exploiting epistemic uncertainty of the deep learning models to generate adversarial samples.

Since we are trying to maximize the loss based on a reference class, we always see an increasing trend during the gradient descent approach of the loss maximization journey.

also in 63,971 other papers