Obfuscation of images via differential privacy: From facial images to general images

Due to the pervasiveness of image capturing devices in every-day life, images individuals are routinely captured. Although this has enabled many benefits, it also infringes on personal privacy. A promising direction research obfuscation facial been work k-same family methods which employ concept k-anonymity from database However, there a number deficiencies that carry over methods, detracting their usefulness practice. In paper, we first outline several these and discuss implications context obfuscation. We then develop framework through obtain formal differentially private guarantee for generative machine learning models. Our approach provides provable privacy is not susceptible outlined produces photo-realistic obfuscated output. addition, demonstrate experimental comparisons our can achieve comparable utility terms preservation useful features images. Furthermore, propose method differential any (i.e., without restriction images) direct modification pixel intensities. addition noise intensities does provide high visual quality obtained via models, offers greater versatility by eliminating need trained model. proposed use exponential mechanism able superior pixel-space using Laplace mechanism.