The Defense Advanced Research Projects Agency (DARPA) are seeking submissions for research concepts in the technical domain of disrupting machine vision systems without in-depth knowledge of how they were built or trained.
This AI Exploration (AIE) Opportunity is called the Techniques for Machine Vision Disruption (TMVD) and is meant to “develop specific techniques to disrupt neural net-based computer vision technology in situations where neither the original training set not the specific vision architecture are available.”
Additionally, the program will help develop image-based attacks that whiten the output vector probability so a successful attack will have a high probability that the output class won’t be the actual class, but without a preference to what output class is actually reported.
For the techniques developed, they must have no knowledge of actual images used to train the computer vision system nor should they have knowledge of the architecture, which includes number of input units, number of hidden layers, number of output units, or specific gradient descent strategy used by the training algorithms during back propagation.
“The resulting technology should be as general or ‘universal’ as possible. This means, first, that the resulting disruption/attack patterns should be effective against a wide variety of images of the subject at different distances, orientations, natural viewing conditions, degrees of occlusion, and image resolution,” the AIE said.