One Pixel Adversarial Attacks via Sketched Programs

Explore a groundbreaking approach to one pixel adversarial attacks on neural networks in this 19-minute video presentation from PLDI 2023. Discover how researchers Tom Yuviler and Dana Drachsler-Cohen from Technion, Israel, leverage program synthesis to significantly reduce the number of queries required for successful attacks. Learn about OPPSLA, a novel synthesizer that employs a stochastic search algorithm to generate customized conditions for adversarial examples. Gain insights into the challenges of one pixel attacks, including the small perturbation region and non-differentiable nature of the perturbation. Understand how this innovative method achieves state-of-the-art success rates with fewer queries compared to existing attacks, and explore its potential for transferability to other classifiers. Delve into the intersection of program synthesis, adversarial attacks, and computer vision in this cutting-edge research presentation sponsored by ACM SIGPLAN.

[PLDI'23] One Pixel Adversarial Attacks via Sketched Programs