[PET] Postdoc position at UniGE (University of Geneva)
claudia.diaz at esat.kuleuven.be
Fri Oct 5 12:49:17 BST 2012
Forwarding in case anyone on this list is interested in applying to this position.
> *Subject:* Reliable content-based access of multimedia repositories
> *Keywords:* multimedia, privacy, outsourcing
> Outsourcing multimedia data is what million of users do everyday:
> people share their images on Picasa, on Flickr, on Facebook or other
> servers. Today, such data is accessed via keywords or tags. Academic
> prototypes already propose content-image search to some extent and, in
> a very near future, these technologies will reach a level allowing
> users to query large image repositories over the Web. This, to some
> extent, already happens with the Google Search by Image service. Such
> services for multimedia data can only flourish as the general trend of
> cloud computing fosters outsourcing.
> These systems create obvious, yet complicated, privacy-preserving
> problems. With the current technologies, it is very easy for an
> outsourced service to identify what a user is interested into, by
> simply checking what queries are submitted to the system. When such
> queries concern sensitive information (medical images for example),
> then, some knowledge can be inferred which might compromise privacy.
> Worse, outsourced services can correlate queries issued by many users
> and subsequently define groups of users sharing some similarity, or
> secretly mine queries to detect patterns and trends.
> The goal of this project is to propose strategies to make more
> difficult such privacy breaches in the context of outsourced
> content-based image retrieval systems. While there are obvious
> anonymization actions to take, more sophisticated techniques are
> likely to be needed.
> Overall, there is a rather long list of possible scenarios that differ
> by the data that is exchanged between the various parties involved
> here, by the level of knowledge each can have, or by the goal which
> might be to preclude the identification of a user and/or the discovery
> of correlations between users.
> Few case-studies can already be foreseen as a context for
> investigating privacy preservation in the content-based image search
> and access. These range from searches based on example query images
> (either sent directly, or as low-level features or even encoded with a
> matched dictionary) to accesses made via browsing and navigation.
> A classical solution consists in sending the actual query with a flow
> of unrelated queries, so as to prevent distinguishing what is of
> interest to the user or not. Determining what the distractors should
> be, their number, their diversity and how to filter the many answers
> received back from the server in order to isolate the true result are
> part of the problems to study. Further questions related to the
> induced increase in complexity along the chain of
> client-network-server need to be answered, in relation to the exact
> scenario that is proposed (eg mobile content access with thin client
> and limited bandwidth). Even further, questions related to whether a
> single user or a group of users is considered should also be answered.
> Similarly, the type of queries (eg very specific rare queries against
> quite generic frequent queries) may be explored. Finally, the level
> of shared knowledge (eg content of the repositories, user expected
> profiles) is a further parameter to be accounted for in the design of
> a privacy preserving strategy.
> Overall, the project can be summarized by investigating what type and
> level of “noise” should be added to the original queries, and how the
> superset of answers should be filtered to recover the targeted
> This work is a tight collaboration between Stéphane Marchant-Maillet
> (U. Geneva, CH) and Ewa Kijak, Teddy Furon, Laurent Amsaleg (IRISA,
> - Postdoc position at UniGE (University of Geneva) asap until end of 2013.
> - Work in Geneva, Switzerland. Exchange with IRISA Rennes, France.
> *Expertise:* information retrieval (search and protocols), data mining
> and machine learning.
> *Application.* Please send your full CV as well as references to:
> - Laurent Amsaleg (laurent.amsaleg at irisa.fr)
> - Stéphane Marchant-Maillet (stephane.marchand-maillet at unige.ch)
More information about the PET