Certified Defense Framework against Patch Attacks on Images

Adversarial perturbation of all the image pixels is computationally intensive and may not be realized in practice. In contrast, an adversarial patch attack where an adversary can choose to perturb a specific subset of pixels in an image, is more practical in fooling a trained image classifier or hiding a person from an object-detection model.
In this project, we study and explore state-of-the-art adversarial patch attacks and their defences. An adversarial patch can easily fool the trained object detector to mark an image as having no object of interest, potentially leading to serious security gaps. Therefore, it is necessary to first identify and highlight images with the possible existence of adversarial patches for human verification

Faculty Supervisor:

Apurva Narayan

Student:

Partner:

TrojAI

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

The University of British Columbia - Okanagan; The University of Western Ontario

Program: