LYNNTECH INC. — Department of Defense SBIR Phase I: N182-127

Description

The Lynntech team proposes to develop a Computer Vision FoolKit system that integrates cutting-edge approaches to systematically evaluate physically realizable adversarial attacks against several leading computer vision classifiers. It has been noted that most deep neural networks are demonstrably vulnerable to adversarial examples, even in the form of small-magnitude changes in intensities of the input images. Adversarial attacks result in an object being mislabeled from the ground truth class or unlabeled (invisible). Mislabeling attacks can target a particular object class label, or be untargeted resulting in a random class assignment. Physical realizations of such attacks on computer vision systems have been a recent active research topic with much debate about their effectiveness and robustness. Our Computer Vision FoolKit will utilize a general gray-box attack algorithm to take into account numerous physical conditions of the images in order to construct robust adversarial modifications to a target’s appearance effective against multiple state-of-the-art classifiers. The FoolKit must incorporate a variety of environmental conditions, including characteristics of images captured from moving platforms, and ultimately extend capabilities across the electromagnetic spectrum. Our FoolKit’s physical attacks will be straight-forward to construct and deploy and effectively fool real-world vehicle and aircraft classifiers in an operational environment.