A Demonstration of Image Obfuscation with Provable Privacy
Liyue Fan · 2019
Widely sharing image data captured by personal or surveillance cameras can enable a variety of research studies, e.g., computer vision, intelligent systems, and social sciences, to name a few. However, image data may contain a range of sensitive information, such as identity, location, and health, which creates significant challenges for sharing image data with untrusted parties, e.g., researchers. Standard obfuscation methods obscure regions-of-interest (ROIs) in image data with pixelization and blurring, which do not offer formal privacy guarantees. Moreover, such obfuscated image data can be re-identified by convolutional neural networks. In this demonstration, we will showcase a novel method for sanitizing sensitive image ROIs with quantifiable privacy guarantees. The audience will observe the obfuscation in a camera live stream, which demonstrates the feasibility of our method for real-time applications. Furthermore, the audience can interact with our method by selecting photos from publicly available datasets and choosing different privacy levels.