Résumé: The incorporation of the mobile crowd in visual sensing provides a significant opportunity to explore and understand uncharted physical places. We investigate the gains and losses of the involvement of the crowd wisdom on users' location privacy in photo crowdsensing. For the negative effects, we design a novel crowdsensing photo location inference model, regardless of the robust location protection techniques, by jointly exploiting the visual representation, correlation, and geoannotation capabilities extracted from the crowd. Compared with existing retrieval-based and model-based location inference techniques, our proposal poses more pernicious threats to location privacy by considering the no-reference-photos situations of crowdsensing. We conduct extensive analyses on the model with four photo datasets and crowdsourcing surveys for geo-annotation. The results indicate that being in a crowd of photos will, unfortunately, increase one's risk to be geo-identified, and highlights that the model can yield a considerable high inference accuracy (48%~70%) and serious privacy exposure (over 80% of users get privacy disclosed) with a small portion of geo-annotated samples. In view of the threats, we further propose an adaptive grouping-based signing model that hides a user's track with the camouflage of a crowd of users. Wherein, ring signature is tailored for crowdsensing to provide indistinguishable while valid identities for every user's submission. We theoretically analyze its adjustable privacy protection capability and develop a prototype to evaluate the effectiveness and performance.

Langue: Anglais