Practical Image Obfuscation with Provable Privacy
Liyue Fan · 2019
An increasing amount of image data is being generated nowadays, thanks to the popularity of surveillance cameras and camera-equipped personal devices. While such image data can be shared widely to enable research studies, it often contains sensitive information, such as individual identities, location indications, etc. Therefore, the image data must be sanitized before sharing with untrusted parties. Current image privacy-enhancing solutions do not offer provable privacy guarantees, or sacrifice utility to achieve the standard ε-differential privacy. In this study, we propose a novel image obfuscation solution based on metric privacy, a rigorous privacy notion generalized from differential privacy. The key advantage of our solution is that our privacy model allows for higher utility by providing indistinguishability based on image visual similarity, compared to the current method with standard differential privacy. Empirical evaluation with real-world datasets demonstrates that our method provides high utility while providing provable privacy guarantees.