This demonstrates the concept of “adversarial examples” from [1] showing how to fool a well-trained CNN. The surprising idea is that one can easily generate examples which the CNN will consistently make the wrong prediction for that a human can easily tell are correct.
[1] Goodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. “Explaining and harnessing adversarial examples.” arXiv preprint arXiv:1412.6572 (2014)