Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:00
huh so good afternoon so i'm severson ourselves so i'm going to
00:00:04
present to our duties in the biometric security and privacy group
00:00:10
ah so the alignment presentation would be a very very quick recap on what is by matrix just in case
00:00:14
you don't know what it is i'm going to talk a little bit about what the user group
00:00:20
what the people working group and what are the different topics on which will working and how just go through a number of
00:00:26
examples of different ways of projects and in time and change into some sort of time or speed up on those ones
00:00:34
right so what is my metrics you probably all know a bit about it so that uh yeah because maybe you have been exposed
00:00:40
to its eyes as to face recognition or fingerprint recognition but you
00:00:44
have also some more exotic ways to the body tricks
00:00:48
um it's about recognition through funds and the use of finding recognition or even think giving recognition
00:00:55
so basically extracting looking at the thing structuring the palm inside the body
00:00:58
basically uh the palm all the finger to authenticate uh recognise it
00:01:04
uh it's a bad fiction broad sense is the automatic recognition of individuals based on
00:01:09
bias characteristics can be biological or um so basically i know jacoby of your
00:01:14
and there is a wide range of applications which is drawing a every and maybe not every day
00:01:20
but now this is a we have a wide uh and it's a um presence of this
00:01:25
sort of application but it's weeks which are not only in um by spot checking all in
00:01:30
video surveillance when you fight you can find even on on in your pocket and before
00:01:35
and there are still plenty uh and you you makes you may think that
00:01:40
by matrix is solved problem because you have it on your life on ten now
00:01:45
which is true will face recognition what they're well on the front end and folded station
00:01:49
now down the challenges all on muscle with the other than senses that would say
00:01:54
um but you have many other challenges which are still remains for instance when face recognition identification
00:02:00
and watch katie for instance uh but also other effects such other problems such as raging
00:02:05
but the moon phase but also by matrix which are and the and the debate
00:02:10
uh and a big big topic which is related to security aspects of biometric data and privacy
00:02:16
aspects of biometric data because by buying three data is personal data and so all this should be um consider wisely
00:02:24
so the group itself um the look at a um a bunch of
00:02:28
topics we don't cover the will feel but at least we focus
00:02:31
on a couple buying take notes such as face speaker in vain uh
00:02:36
looking at recognition from from different point of view different senses
00:02:39
a speaker uh so that you know there is an overlap with the speech group has met you presented this morning swept on projects
00:02:46
bring recognition so we have been looking at the using local feed zero g. in
00:02:49
the past the c. g. o. u. g. to authenticate sky plus people
00:02:53
the use of sauce by matrix so using the same data but do do all these
00:02:57
things then recognising people which often snow estimating uh the age or recognising the gender
00:03:03
uh you can do also cool things such as trying to estimate the heartbeat
00:03:08
from of a phase you do so i'm going to have an an example project like that
00:03:13
another big topic is uh related to what we call spoofing all presentation attack now nowadays
00:03:19
um we have a this is a big topic for us well looking
00:03:23
at to doing even yeah but you can use is about existence
00:03:26
uh against with this this presentation of facts to one example of those attacks in the back
00:03:32
um and therefore while working on how to detect those attacks
00:03:37
so working on presentation up and detection but also
00:03:41
and that includes morphing and did fakes detection from some of you that have it about the fakes
00:03:48
uh recently so i was we're going to have an example on that uh we start to
00:03:52
look at also tempered protection and a web be also a number of activities on
00:03:56
all to combine all this together and finally we have a big focus and all work
00:04:01
is on the people just the research actually it's almost a priority for for us
00:04:06
so we try to do whenever possible uh to uh
00:04:11
to to systematically publish cooled opens will spend
00:04:14
data and documentation to make the work they punishable as much as possible as easy as possible
00:04:23
right so the team is composed of one permanent which is a which
00:04:27
is myself and uh so we had the previous a research
00:04:30
associate on ground rules which now has been hired as a permanent so it's going to um the setting up a new group
00:04:36
so is facing out from the team now so but that now the
00:04:39
group is composed to fall research associates um of six pots
00:04:44
and does free p. h. d. students so when one of the p. h. d. students actually because
00:04:48
provides just by someone from the speech group so much you actually and uh also an intern
00:04:55
uh in terms of outputs from the group is also figures are represented the two years ago
00:05:01
or two during that that even so it what it will do a little bit
00:05:04
so i'm still a production from the german conference papers now we have a
00:05:08
second book actually this is a set a edition of the previous book
00:05:12
because we launch a series of book on politics pushing an anti spoofing
00:05:17
so now we are running the second edition of that books
00:05:21
and um we also publishing this written quite a number of probably
00:05:25
data sets will graduate students from e. p. f. l. mostly
00:05:28
we now we teach other taught school at e. p. f. l. must of course level add to list of laws and you pay for
00:05:35
and we also involve many technology transfer activities because of the domain as you can imagine is very and trendy
00:05:42
uh so um a lot of a lot of uh projects with the industry
00:05:47
number of patents to speed of companies from the t. v. input into the man that
00:05:50
was mentioned by over this morning just enough on the group that was acquired by
00:05:55
a messy d. um uh in enjoy in austria uh early this year
00:06:01
write a letter number of patients that graduated from e. p. f. l.
00:06:06
a number of master students that and find jobs by spaces messages in companies will wind but of
00:06:13
collaborations findings could maybe community now since and so if we about ten million swiss francs
00:06:20
uh from virus funding agencies a swiss city high
00:06:25
master of course uh you you no way no way we have product would know norway for instance funny but no
00:06:32
knowledge and and such urgency and of course i also uh i uh that i know yeah as
00:06:37
a mention by yeah this morning what projects with us and companies that do for you
00:06:43
right so i'm just going to go through um for five gnome project situation very quickly
00:06:48
so one of the first project example is uh on it on a genius face recognition so this is the
00:06:54
work of a p. h. d. students so check which may be in the office in the room today
00:06:58
so um where the the goal here is to try to match feces across the full spectrum
00:07:04
right so basically uh how do you solve the problem of because of the problem
00:07:08
where if you want to bet match the other picturing your spectra and another
00:07:11
one in its in unified optimal or even the points is catch up you
00:07:14
matches the two together so um we have a approach it we've uh
00:07:20
we've basically convolutional neural networks uh uh and actually one of the things what well what
00:07:25
certain we try to do all we up i put the size is that
00:07:28
basically uh the very first layers which are actually the low level a label features
00:07:34
of the of the of the recognition process actually domain domain dependent and
00:07:39
then we demonstrated empirically that by just taking a three train
00:07:44
a network for visual spec trough and train for face recognition i just adapting the
00:07:49
very few layers which makes uh actually if you if you hundreds of parameters
00:07:54
a few thousand to find it on the first day as you can use
00:07:57
the same network missing architecture to match faces from the phone spectrum
00:08:02
then we have shown basically by stealing all all all deep you adapt
00:08:07
the network that just adapting the first three years are enough
00:08:11
containing a white but is is that the low level features which on
00:08:15
the top at the beginning that say all domain specific and want
00:08:18
you all those on the menu dependence basically they do their job
00:08:20
of separating between identities but the first one normal domain dependent
00:08:25
some with this this isn't this is work and uh and the uh the situation in the region actually
00:08:31
the second piece of work here relates to the these reality to to spoofing
00:08:35
to presentation of facts does it like this example in the back
00:08:39
and here um this is actually the second this is the extension of a previous work
00:08:43
where we demonstrated with uh basically i'm a set of the yard face recognition system basin convolutional
00:08:49
neural networks they wanted to demonstrate if they will view hubble as they are actually white
00:08:54
what you didn't demonstrate that they also very even other tracks so we didn't verify this i put
00:08:59
this is to traditional attacks made of paper read that text the screens and so on
00:09:03
they wanted to get to see if it was possible to do it with very hard against attacks that now very
00:09:08
expensive but might be more challenging the future when more or less expense in the future more particularly silicon asks
00:09:16
so uh what the question we had is how did face wouldn't systems operating either
00:09:21
so what we did is that we demonstrated indeed by making it at a quite a
00:09:26
bit so we built we work we manufacture to build custom c. can masks
00:09:31
so you may you have to keep in mind that this study was the now we have seventeen of business
00:09:36
each of these mass cost about three paulsen mules so three zero zero zero three thousand you know
00:09:44
so it's a very expensive but check to make slides but it's also require it was required
00:09:50
so we make a study that we've um a number of different uh identities and asks
00:09:55
so it's it's a result of quite a long project because you can imagine to find this money to
00:10:01
to complete the study into manufacture the mask to took nearly a year on the masts it wasn't very long study and so
00:10:08
now we have been able to demonstrate the reality of the face recognition to the for these these these kind of messed
00:10:16
right so it's just an example and we have been able to show that the
00:10:18
variability here made measured in our shop will as a p. m. r.
00:10:24
is ten times higher than the normal force much rate which would be the normal force much
00:10:28
rate by usable for people still so and and and and uh impulse to visit
00:10:36
and there's a form of variability which is more recent so this is a work from a
00:10:41
parallel question of in the context of um of a dab of that happening projects
00:10:46
we wanted to measure the reality or face recognition systems to be fixed
00:10:51
so did face maybe some of you other updates this technique that come from machine owning it
00:10:55
even where you are using the ends mostly so you not to notice and modest too
00:11:00
converts to map the conductor face image into the face image of someone else
00:11:04
and then you swap it into into video so you can use it in in different ways
00:11:08
and here the in the example to read your we have and to make
00:11:12
this female subjects into major subjects any other what you can see
00:11:16
is basically we to that subject a and b. and and this is actually subject uh
00:11:22
eight in which we have inserted the face of subject sorry this is in image
00:11:26
of subject to be which was inserted this on the face of subject it
00:11:31
and this is the oppose its decision image actually off subject eight where we haven't set to the face was not
00:11:38
the what fate the the face to convert it into that things so we do do well database of
00:11:45
such swaps to make experiments where i find that in our eat your face recognition and it did
00:11:52
we uh so this is just published is going to just soon and we'll be able to show the reality
00:11:57
again that to two of the most to mostly use description systems of any other you you've been able to be
00:12:03
fakes and it's even more when you have a high quality um high quality fix windows options are you
00:12:11
this is an example in the video so you're again this is an original source that you want to
00:12:16
temper that you want to modify to basically want when sets the phase of that person into that
00:12:21
that video so we take a eyes another video that person here and we said that face into that you do so this is the result
00:12:29
and then you take that video that okay every frame of the video when you plug it into face recognition system
00:12:34
and you are going to observe basically some blue this interface we can score and other times mostly
00:12:40
above the threshold when these green showing that most of the time it's really it's
00:12:43
exceed the matching threshold to it would be organised as this person right
00:12:49
and so yeah this is not the output of the nozzle reason that is able to detect when this it if they cannot
00:12:57
another example would show also relates to on price proofing so this time it's is the
00:13:02
on this work is done the context of and i have been funded project
00:13:06
uh where we have to look at the wide range of possible of attacks with out already
00:13:11
twenty and crazy attacks such as using fake it's um
00:13:14
the sound level of masks rigid silicon plastic
00:13:18
vision glasses paper prints on white very including makeup as well
00:13:23
and the it has to be robust to environmental conditions
00:13:26
and there was as it was i apply funded project there is a target in terms of performance that you
00:13:30
have to reach me because if you don't reach it to go out of the project basically so um
00:13:36
yeah the challenges that you have to meet this target buttons you have a limited amount of data
00:13:41
you would be to do you know how nice to uh all this from a fedex
00:13:44
so one of the recent word was um that was proposed by someone from the team to enjoy it
00:13:49
is to use uh the multi put the multiple channels all because we recorded at the site we have um
00:13:54
different uh spectroscopy respect pollen you confide them and dips as well
00:14:00
and then combining all this information together we've uh c. n. n. actually that it's a deep c. n. n.
00:14:05
which is which is multichannel retrained visual spectra and then just
00:14:10
again adapted on the first layer for the different domains
00:14:13
we have shown that we can outperform basically on on on we cannot perform the different channels
00:14:20
uh by far and even whole ball foam uh this confusion for the dish on the different channels
00:14:26
so the this work is under profession for to be to be sedated is it on
00:14:31
last word each its own copy p. g. so remote photo pretty small roughly
00:14:35
this is a topic which is it is not for couple of years of a difficult topic
00:14:39
because there is a actually a lack of to be databases and a lack of
00:14:44
open source implementations for that so there is a lot of claim this domain but very little uh guarantees
00:14:49
that yeah actually manager but they can reach so one off we spent nearly more than a year
00:14:56
put it in place baselines to try to reproduce work from was us and we failed
00:15:02
so we fail because the data is not available because the code that is not available when the code is available
00:15:07
in shared with us it's it's incomplete any blacks plenty of
00:15:11
little details in tricks that are not written anywhere
00:15:15
so nevertheless we uh we observe that we could use this signal
00:15:20
that is sense basically by just looking at some supper evaluations of the skin colour
00:15:24
or in the in the video you can estimates so do you have here
00:15:29
in red the so read the signal which is estimated to happy to just ignore any
00:15:35
green is the act real v. p. so the blood volume a wooden posts from
00:15:40
a sensor which is a fingertip by that you actually measure so this is the brunt rules that say they want us to make this as close as possible
00:15:49
so this is actually arranging to be able to estimate that's personal
00:15:52
spree it very accurately but nevertheless we manage to use it
00:15:56
to detect fakes so when run experiments where we will have been a friday
00:16:02
subject so we'll be people while using the doppler signals and then
00:16:07
we managed to basically user class for you to do that and i some
00:16:11
statistics on the on the on the after spectral analysis of the signal
00:16:15
to uh or to run a classifier to detect the to find difference between real faces and masks because obviously
00:16:22
masks don't have basically don't have don't don't have balls i mean this is this
00:16:26
is this is this an on an object that doesn't have a that they
00:16:30
don't have the same the colour or the coalition than the real face so you
00:16:34
should be able to find uppers which in fact this is wrong because
00:16:38
this is another go isn't that estimates uppers a signal from the video so
00:16:42
it does the job means is it is going to generate the plausible
00:16:46
a post rate but still there is there is some discriminant information
00:16:50
to a low separating the real faces from that next
00:16:55
so and with me that we have been able to show that we owe it still passing goals are up
00:17:00
p. p. g. based uh aunt is pushing the um um other reasons on four different data sets
00:17:06
and we are on that show also probably work we are sharing the data
00:17:11
and we are sharing the source code falls asleep with his or

Share this talk: 


Conference program

Introduction by Hervé Bourlard
BOURLARD, Hervé, Idiap Director, EPFL Full Professor
29 Aug. 2018 · 9:03 a.m.
Presentation of the «Speech & Audio Processing» research group
MAGIMAI DOSS, Mathew, Idiap Senior Researcher
29 Aug. 2018 · 9:22 a.m.
Presentation of the «Robot Learning & Interaction» research group
CALINON, Sylvain, Idiap Senior Researcher
29 Aug. 2018 · 9:43 a.m.
Presentation of the «Machine Learning» research group
FLEURET, François, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
29 Aug. 2018 · 10:04 a.m.
Presentation of the «Uncertainty Quantification and Optimal Design» research group
GINSBOURGER, David, Idiap Senior Researcher, Bern Titular Professor
29 Aug. 2018 · 11:05 a.m.
Presentation of the «Perception and Activity Understanding» research group
ODOBEZ, Jean-Marc, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
29 Aug. 2018 · 11:24 a.m.
Presentation of the «Computational Bioimaging» research group
LIEBLING, Michael, Idiap Senior Researcher, UC Santa Barbara Adjunct Professor
29 Aug. 2018 · 11:45 a.m.
Presentation of the «Natural Language Understanding» research group
HENDERSON, James, Idiap Senior Researcher
29 Aug. 2018 · 2:03 p.m.
Presentation of the «Biometrics Security and Privacy» research group
MARCEL, Sébastien, Idiap Senior Researcher
29 Aug. 2018 · 2:19 p.m.
Presentation of the «Biosignal Processing» research group
RABELLO DOS ANJOS, André, Idiap Researcher
29 Aug. 2018 · 2:43 p.m.
Presentation of the «Social Computing» research group
GATICA-PEREZ, Daniel, Idiap Senior Researcher, EPFL Adjunct Professor
29 Aug. 2018 · 2:59 p.m.

Recommended talks