Authors: Xavier Alameda-Pineda, Miriam Redi
Affiliations: INRIA Grenoble Rhône-Alpes, Wikimedia Foundation
For this edition of the SIGMM Opinion Column, we carefully selected the discussion’s main topic, looking for an appealing and urgent problem arising for our community. Given the recent Cambridge Analytica’s scandal, and the upcoming enforcement of the General Data Protection Act in EU countries, we thought we should have a collective reflection on ‘privacy and multimedia’.
The discussion: multimedia data is affected by new forms of privacy threats, let’s learn, protect, and engage our users.
Users share their data often unintentionally. One could indeed observe a diffuse sense of surprise and anger following the data leaks from Cambridge Analytica. As mentioned in a recent blog post from one of the participants, so far, large-scale data leaks have mainly affected private textual and social data of social media users. However, images and videos also contain private user information. There was a general consensus that it is time for our community to start thinking about how to protect private visual and multimedia data.
It was noted that computer vision technologies are now able to infer sensitive information from images (see, for example, a recent work on sexual orientation detection from social media profile pictures). However few technologies exist that defend users against automatic inference of private information from their visual data. We will need to design protection techniques to ensure users’ privacy protection for images as well, beyond simple face de-identification. We might also want users to engage and have fun with image privacy preserving tools, and this is the aim of the Pixel Privacy project.
But in multimedia, we go beyond image analysis. By nature, as multimedia researchers, we combine different sources of information to design better media retrieval or content serving technologies, or to ‘get more than the sum of parts’. While this is what makes our research so special, in the discussion participants noted that multimodal approaches might also generate new forms of privacy threats. Each individual source of data comes with its own privacy dimension, and we should be careful about the multiple privacy breaches we generate by analyzing each modality. At the same time, by combining different medias and their privacy dimensions, and performing massive inference on the global multimodal knowledge, we might also be generating new forms of threats to user privacy that individual stream don’t have.
Finally, we should also inform users about these new potential threats: as experts who are doing ‘awesome cutting-edge work’, we also have a responsibility to make sure people know what the potential consequences are.
A note on the new format, the response rate, and a call for suggestions!
This quarter, we experimented with a new, slimmer format, hoping to reach out to more members of the community, beyond Facebook subscribers.
We extended the outreach beyond Facebook: we used the SIGMM Linkedin group for our discussion, and we directly contacted senior community members. To engage community members with limited time for long debates, we also lightened the format, asking anyone who is interested in giving us their opinion on the topic to send us or share with the group a one-liner reflecting their view on privacy on multimedia.
Despite the new format, we received a limited number of replies. We will keep trying new formats. Our aim is to generate fruitful discussions, and gather opinions on crucial problems in a bottom-up fashion. We hope, edition after edition, to get better at giving voice to more and more members of the Multimedia Community.
We are happy to hear your thoughts on how to improve, so please reach out to us!