Targeted adversarial attacks with Keras and TensorFlow

Description

Summary

• Accepting an input image and determining its class label using a pre-trained CNN Step #2: Figure 1 of Explaining and Harnessing Adversarial Examples) The gist is that adversarial images are purposely constructed to fool pre-trained models.
• Here, we input the image of a “pig” — the adversarial attack algorithm then perturbs the input image such that it’s misclassified as a “wombat”, but again, we did not specify what the target class label should be (and frankly, the untargeted algorithm doesn’t care, as long as the input image is now incorrectly classified).
• Reader Interactions Before you leave a comment... Hey, Adrian here, author of the PyImageSearch blog.