Python: Advanced Guide to Artificial Intelligence

Adversarial examples

The image classification algorithms have reached human-level accuracy on several datasets. But they can be easily fooled by adversarial examples. Adversarial examples are synthetic images that fool a model to produce the outcome that is needed. Take any image and choose a random target class that is incorrect. This image can be modified with noise until the network is fooled as show by Goodfellow et al. (https://arxiv.org/pdf/1412.6572.pdf). An example of an adversarial attack on the model is shown here:

Reproduced with permission from Goodfellow et al.

In this figure, an image is shown on the left with 58% confidence ...