This project accompanies the lecture deep learning (392221; Module 39-M-Inf-DL) and handles the GTSRB - German Traffic Sign Recognition Benchmark.
A classifier from an existing network has been used for applying several different attacks (few pixel attack, fast gradient, backdoor poisoning, universal perturbation). The success rates of the different attack methods are discussed (measures: accuracy and average perturbation) and evaluated concerning their noticability by a human.
The Mischief-Makers Ferdinand Stoye, Niclas Kopp and Bianca Schröder are proud to present!