Room 0120 Green Hall Rodin Auditorium
As vision algorithms become increasingly successful "in the wild," many modern machines and devices come equipped with cameras for visual sensing. Thus, an increasing number of images captured by cameras are meant for consumption by algorithms rather than directly by humans. This raises the question of whether standard acquisition with a traditional digital camera is the best way to capture information that is most useful to these algorithms. At the same time, given the versatile utility of visual information, it brings up the possibility that the images captured by these on-device cameras could also be used by an adversary to recover sensitive information about the device's surroundings. In this talk, I will discuss two recent works: the first is a learning-based approach for optimizing the acquisition pipeline of a camera so that the information it captures is most useful for subsequent computational processing. The second discusses an approach to find an encoding function for images such that the encoded images permit a variety of useful inference tasks, but inhibit the recovery of designated sensitive attributes.