Computer Vision News - December 2019

be happening. One of the things is that the future activations in these networks are not spatially invariant even without the attack. You see some activations even when the activations should be zero. The other problems are that in these pyramid networks you see that they predict flow even when there is no motion and even when there is no attack. What happens during the attack is these factors get amplified and the predictions become really random.” In the real world, Anurag tells us this amounts to vulnerabilities in several applications where optical flow is used. For example, self-driving cars use optical flow to compute motion . You could have a patch on the street and the optical flow predictions could be really bad . In that case, the decision taken by that self-driving car would be incorrect. He is modest when we suggest that this could be life-saving work: “It could be. It’s promising, yes.” Anurag says their work has given them several ideas about what these attacks can do. In future, they will be exploring other ways to attack these systems to reveal problems with the networks. They can then propose defenses to make the systems more robust so that they can be deployed safely in the real world. "but we proposed some heuristics of what might be happening." Anurag Ranjan 13 Best of ICCV 2019