Billions images are shared daily on social networks.
When shared with an inappropriate audience, user-generated images can, however, compromise users' privacy and may have severe consequences, such as dismissals. To address this issue, different solutions were proposed, ranging from graphical user interfaces to Deep Learning (DL) models to alert users based on image sensitivity prediction. Although these models show promising results, they are evaluated on datasets relying on small participants' samples.
To address this limitation, we first introduce SensitivAlert, a dataset that re-annotates the previously annotated images from two existing datasets, but using a German-speaking cohort of 907 participants.
We then leverage it to classify images according to two sensitivity classes---private or public---using recent transformer-based DL models.
In our evaluation, we first consider consensus-based generic models using our dataset as benchmark based on image content itself and its associated user tags. Moreover, we show that our fine-tuned models trained on our dataset better reflect users' image privacy conceptions.
We finally focus on individual user's privacy estimation by investigating three approaches: (1) a generic approach based on participants' consensus for fine-tuning, (2) a user-wise approach based on user's privacy preferences only, and (3) a hybrid approach that combines individual preferences with consensus-based preferences. Our results finally show that the generic and hybrid approaches outperform the user-wise one for most users, thus ensuring the feasibility of image privacy prediction preferences at the individuals' level.