CVPR Daily - Thursday

The solution proposed here is called output code matching. It is drawn from communications theory, implementing something similar to error- correction codes in communications systems to render deep neural networks more robust to these attacks. This encoding dramatically reduces attack stealthiness , showing that this uncomplicated method effectively defends against the problem of stealthy weight bit-flip attacks. Ozan would like to acknowledge two early collaborators on this work, hardware fault-injection experts Jakub Breier and Xiaolu Hou , who introduced them to the problem at a seminar. “ Those discussions were a nice kickstart to get this work on our table and sparked us to look at it from the other side, ” he tells us. Ozan views CVPR as a great way to learn about new problems like this one and ensure the community sees their progression in subsequent years. 28 DAILY CVPR Thursday Oral Presentation “ When you discover a novel problem at CVPR, you have a whole year to tackle it, ” he points out. “ You can develop a simple new approach or alternative solutions, and you take away something that’s a problem and make it possible. ” To learn more about this work [ID 9002], come to oral session 3.2.1 today at 13:30 and poster session 3.2.