Adversarial Examples and Uncertainty

While neural networks can classify images with very high accuracy, it was shown in 2013 (original paper by Szegedy et al) that it is also possible to make very small perturbation to an image so that the network misclassifies it (e.g. so that a panda is classified as an airplane). Many variations of this effect have subsequently been discovered and studied, but the mechanisms underlying this are still not well-understood. In this project, we plan to look at the systemâs underlying uncertainties and certainties that might contribute to this effect. For example, even though the system might predict that the panda is an airplane with probability of 98%, is there some other way in which the system is in fact uncertain about this? How might we be able to quantify this?

Faculty Supervisor:

Evangelos Milios

Student:

Chandramouli Sastry

Partner:

Vector Institute

Discipline:

Computer science

Sector:

University:

Program:

Accelerate