Transcriptions
Note: this content has been automatically generated.
00:00:00
yeah i see alright hand that you sell out of it i think it
00:00:06
would be really interesting tool so i have a question about secure multiparty computation
00:00:11
a long wrong but the way i understand it boards um is that it's based on the assumption that the different parties
00:00:18
will never clued yeah because of all the parties can do
00:00:21
they can they can uh uh quite a description case saints case
00:00:25
so well and tractors how do you decide what is
00:00:28
an appropriate number of parties such that they want clued
00:00:34
yeah so so i'll in the approach we we design we
00:00:37
use the the so called any trust the right model were
00:00:42
all parties but one can conclude if you have one but doesn't colour the system is secure
00:00:49
so the more parties you have the more purely the system is more secure
00:00:53
but how to get to the exact number in fact it's like i'm talking say
00:00:58
you know like fine for like i was on any particular cases you you just need it's you just need and minus one
00:01:06
so and minus one can conclude and and one one uh you see what i mean
00:01:11
to threshold on and what is the minimum number of parties that don't have to go to
00:01:16
these guys would be one i understand mathematically but um what i
00:01:20
don't understand and practised how do you go about deciding on a number
00:01:25
and not and number just um like um
00:01:29
i mean even if it's an minus one right yeah i mean still check of the situation what does doesn't hold on tractors
00:01:38
have some ten yeah i think in the end there's a there's always some legal protection around
00:01:43
that so that's what what happen in reality we're uh we have to make contracts were people
00:01:50
agree to knock on wood somehow right into not sure if
00:01:53
you really don't okay uh that's oh so technically the system
00:01:59
preserve security as long as there's one party that that does include with the others
00:02:04
even if all the others are compromised uh but practically uh how do we ensure this is through legal code
00:02:11
okay and then in your applications with the whole school how many parties would use typically
00:02:16
uh in the application i showed you the it's it's the five universe documents with
00:02:20
so the rebar little bows and geneva so five party answer yeah five okay yeah
00:02:26
but this is it's i mean theoretically it's unlimited rights really okay
00:02:29
at as many times as you want yeah okay they give him
00:02:37
um but i must say it's yes
00:02:44
and how beautiful presentation i just so one question um about the
00:02:49
slide you showed of for the um the that most of them
00:02:54
for the the the production information comes or switzerland uh is the solution does this one that you use
00:03:01
a solution uh accept to the to the you all
00:03:04
read it so o. e. does your review of the
00:03:09
the two solutions to go no i think it does kind of the feedback
00:03:14
you gave on the approach right we're to essentially the the whole trick was to
00:03:19
to never have the the corruption gene pool uh_huh uh so
00:03:23
this was basically guaranteeing that there's no way of decrypt doing
00:03:27
unless all these editions involved in the process agreed to to
00:03:31
do that right so as long as they these under interruption
00:03:34
this can be consider anonymous so that what is that what is it back uh_huh and this was instrumental also too
00:03:41
a move forward also with canton or um
00:03:45
protection authorities and ethics
00:03:47
come up that you can
00:03:55
any other questions that nice i have one question still yes i'm new
00:04:01
here see if we have a lot of people working on has a ah
00:04:05
keychain so we i i often hear people around saying oh
00:04:09
yes meeting tonight and there's not enough data unless it yes it's
00:04:13
a problem but i was wondering if this privacy in your field
00:04:16
one of the main issues are they also other factors that uh
00:04:20
need to gauge and not being shane as much and yeah domain as i mentioned before i think there is a
00:04:27
that more of a challenge that we have to change like this data ownership feeling right so we're done
00:04:33
there's a lot of back at it academic competitions right so who on b.
00:04:37
b. s. everyone just want to exhaust any possible question before you can have how
00:04:42
and uh so so the and this is a problem right so uh i think
00:04:48
uh and that doesn't have to be good price so so i
00:04:52
think this is one problem and also because of their decision especially
00:04:57
seen in uh in health care as i mentioned before like they are
00:05:02
like we award data is collected through proprietary systems than
00:05:07
and every property system as his own way of storing and modelling uh the uh the semantics of the data
00:05:13
so there's a lot of effort that has to be put before you can put any federated system place right on mapping the
00:05:20
that constant yeah and they're they're huge initiatives in europe uh
00:05:25
to to to map these to a so called common become models
00:05:28
um to to you know really try to sort of the the
00:05:32
pro right uh_huh it's a lot of manual manual work into that
00:05:38
and a lot of data that is out there that we're not using other because it's that's exactly
00:05:42
exactly yes especially if you go also have outside if you have i guess he that to other countries
00:05:49
less developed countries also i think yeah even more states i think
00:05:52
it's where light like cadillacs could have a great impact liking standardisation
00:05:58
uh i think are the for example structuring yeah unstructured data
00:06:03
and try to might be up to two semantic standards this this is something uh_huh
00:06:08
useful and not that awful the and interesting and yeah
00:06:15
ah yes okay thank you very much for having many interesting topics
00:06:22
and of course you can delve into the details and different
00:06:25
technologies at a poem controlled by i have more say pragmatic
00:06:30
uh yeah action action hospitals in switzerland okay definitely be confined
00:06:34
in agreement we can find the third trustee party and we
00:06:37
can provide data and these guys can train in the commonplace
00:06:40
unless you think about some i would say that ownership or
00:06:45
whatever commercial but in make okay so it's it's let let let let's just need to get these discussion out of this but
00:06:51
i'm more concerned about the the phone aspects get talking about
00:06:54
performance data instead of the different hospitals like will but different equipment
00:06:58
different equipment is categorise the like different resolution like different sensitivity yeah
00:07:04
by different parameters and its concerns
00:07:06
everything from microscope to whatever imaging device
00:07:11
and here you might think okay the protected data incidence of that um but you can clearly understand that
00:07:16
the the train the models on the different take originating from different sensors say let's call it looks at
00:07:22
you might have much more she'd use concerns in their need of shrink his data
00:07:28
rather than just protecting privacy and your simple standardisation probably format of the
00:07:32
data will not help much if the data is a lot different physical nature
00:07:37
yeah and here my mean my my personal doubt when i started to sink a about all this application
00:07:42
it's great it's cool it's still have it but basically even hospitals would be the you you provide them privacy
00:07:48
protection protocols but whether it's the useful and practical and can put magically
00:07:53
practical that uh if we use unit you should will converge to the data
00:07:57
we get from different sensors yeah no i think i agree with you so
00:08:03
i i think uh this doesn't solve the problem so uh it it's uh
00:08:09
i think what what will be it's it's a complimentary solution if you want to
00:08:14
to centralise india i think's improvisation will not disappear a bigger country but not disappear because
00:08:20
as you mentioned there are situations in which further it under
00:08:24
any would cause more harm in terms of introducing bias or
00:08:28
uh uh and things that you cannot leave and monica or because you don't have access to the original data
00:08:34
uh what i what i think is that it could be
00:08:37
a way to facilitate some expose initial exploration of beta before central
00:08:43
position or or caring go out uh uh you know simple
00:08:48
analysis or training maturity models where you are pretty sure about standardisation
00:08:53
um but of course is not is not the do both solution that will
00:08:58
um you know substitute to traditional beta sure i think i think the two approachable
00:09:04
quite this is just that we if we want to driving to centralise data sharing
00:09:10
especially across borders the stakes really really mark uh in the eh
00:09:14
speech and i i shortly before they this whispers eyes at network
00:09:18
which is currently adopting morse improvisation on trusted research environments like which a lot of
00:09:25
uh access control security even with that it takes one year for people to agree
00:09:30
uh to to to sign contracts and so on so imagine these at the international scale
00:09:35
so i think federated learning at federated analogies can play a role in sort
00:09:40
of accelerating the first access to the data but of course it is not final