Meeting minutes
debbie: explains interoperability
hugues: this is dangerous because your personal assistant knows a lot of information that is private
… I don't know how to teach a machine the different layers of privacy
gerard: need to anonymize the question
hugues: I love the idea
hugues: I am developing two things. I am developing a simple, rule-based dialog system in the interests of getting something to work rapidly
… when I get the answer from the person I will analyze it in an embedded AI
… from that I build a Knowledge Graph of the person, then I can query the KG about the person to build the life of the person to be able to write my bio
… you can tell things to the ghost writer that you don't want to be disclosed to certain people. The human ghost writer can understand that but I don't know how to tell that to the machine.
… if I give my companion my credit card PIN, it can buy something, but I don't want anyone else to get my PIN
… people are not supposed to access my companion without authorization, or understand different levels of privacy
… you can't be sure that people will respect different levels of confidence
gerard: we have to trust our major-dome "personal assistant"
… like Knowledge Navigator
hugues: how to define rules to keep information secure
… the KG can mark information as private by adding links to each vertex to indicate level of privacy
… from 1-10
… the issue is in a dialog not everything is private, and it's hard to say "this is private"
… the issue is that you might develop too much confidence in the companion
… this happens with humans, you might say things that you forget are private.
debbie: link?
hugues: no, but arxiv paper to be presented next week
gerard: presented this work at Conversational Interaction two years ago with Speech Morphing
hugues: what is new is that now we are taking advantage of embedded AI
… now we are asking for weather, Wikipedia information
… my own data are not public
… I've been thinking about this kind of interaction for a long time
… there is semantics attached to privacy, for example, I know some things in the defense industry, but they don't want to talk about at home
https://
hugues: this is a very complex subject
… for example, in the defense industry, they compartementalize information
… this is a recurring question
https://
debbie: two problems how to designate private information in the KG, and how to tell companion that some information is private
hugues: should companion ask the user if information is private
debbie: sometimes you want the companion to buy things for you
… for example, Alexa can buy things for you
… could you just ask the user to tell the companion to keep certain information private?
hugues: yes, within a session
… there is a big problem with the definition of privacy
debbie: a lot of people want to tell you about AI and privacy, but I don't know if they have a clear idea
hugues: you can keep all information in a closed box
hugues: the simple fact that your companion connects to something means something
… I'm now working on a closed system, that doesn't connect to the cloud, everything is kept inside the box
… mostly I query the system myself
… from that information I can build a biography of the user, when the system writes a biography of the person certain information needs to be kept out
… some information can be provided to my wife, children
dirk: do you label the information?
hugues: the different kinds of information are recorded in the KG
… in a graph you have many roads to a certain point, you might get to the information from an unprotected path
… the user has to say "this information has to be protected"
… I used to work for the defense industry, we knew that everything in a room has to be confidential, and it's still confidential after we leave the room
… how can a machine understand that?
https://
dirk: also have to consider trusted environment
hugues: in a trusted system we can make rooms so that the system cannot disclose certain things
hugues: say I want to buy trousers, the seller will recommend trousers based on my skin color, eye color, information that I would like to keep confidential
dirk: you can derive confidential information from other information that you already know
… privacy is a lot more than on/off
… very interesting work
https://
https://
debbie: we should take a few weeks to take a look at these papers and invite hugues back
dirk: how can we make use of this work?
hugues: it is not because we don't have the answers that we shouldn't do the work
… it is a matter of semantics
debbie: can send hugues the link to subscribe
hugues: this is the most important subject in AI