serre-lab/Adversarial-Alignment

Scaling-up deep neural networks to improve their performance on ImageNet makes them more tolerant to adversarial attacks, but successful attacks on these models are misaligned with human perception.

Jupyter NotebookMIT