Hardware Accelerator for Adversarial Attacks on Deep Learning Neural Networks

By “attack” the authors mean figuring out ways to trick the network into classifying things incorrectly. For example, you might attack a network that recognizes faces by determining certain inputs that still are recognizable as faces to a human, but not to the network.

Yes: you can (for example) confuse a vision algorithm into thinking an image of a turtle is a rifle https://mashable.com/2017/11/02/mit-researchers-fool-google-...

Yeah, many ways. Besides what the article describes there are at least 5-10 known attacks at the moment.