Talks

Semantic Adversarial Perturbations using Learnt Representations

Speaker

Isaac Dunn

Abstract

Adversarial examples that fool image classifiers are typically created by searching for a small norm-constrained perturbation to the pixels of an image. However, it is increasingly acknowledged that such perturbations represent only a small and rather contrived subset of possible adversarial inputs. I will present a novel method for the construction of a rich new class of semantic adversarial examples; by leveraging the feature representations learnt by generative models, this approach makes adversarial but realistic changes at different levels of feature granularity. For instance, the attack is able to perturb the pose, location, size, shape, colour and texture of the objects in an image. Significantly, these abilities are learnt rather than implemented in ad-hoc ways.