People are using ConvNets all  over the companies. The adversarial examples of one domain  have the capability to transfer into another domain. And with the  Convnets portability i.e the ability to copy the weeks or months trained convnets and use it for their use case, combined with the possibility  by someone to accidentally or intentionally  transfer the adversial examples into our domain and screw the ConvNets over, these adversarial networks pose  a challenge to the  ConvNets.

For example adding a  slight ostrich weights  to a dog picture will cause the convnets to mistakenly identify the image object. i.e

adversial networks.jpg

Specific Serious Implication Example: Add adversial example to face recognition and the face recognition system is likely screwed, mistakenly identifying other people.

Advertisements