Ninja Codes: Exploring Neural Generationof Discreet Visual Codes

Abstract

In this paper we report the results of our early explorations regarding Ninja Codes, a new class of visual codes intended to be used in a variety of interactive applications including augmented reality, motion/gesture control, contactless data transfer, robotics, etc. By harnessing the power of adversarial examples, Ninja Codes can be rendered discreet, concealed to human eyes but easily recognizable to detectors based on deep neural networks. The paper will provide a high-level overview of Ninja Codes, and describe an initial, proof-of-concept implementation built on top of existing face detection software. We see this work as a promising step toward a new family of methods by which digital information can be seamlessly encoded into real-world objects and environments.