i

so my name is patrick bob um from france of a little uh uh and also france

and so this talk is gonna be a

uh what i'm knocking security a mostly out to design secure well uh and baiting scheme

and then will show you how to use a nice mess medic zero record optimal to once per in order

to minimize the distortion and and also to guarantee what um

security

um

so the line is very simple a first introduction on the

and what are mocking security is and uh i will uh

propose requirement in order to achieve secure and baiting

uh i would propose so different embedding scheme and compare them using a performance and that is this

and then will conclude my work

after the what

okay uh what's mean security in what are marking so a security as thing to do is do you picked

compression uh no what at to share additional like

go and uh

channel

eight

uh you have to consider a first and that this ad that's every

it's is very important because the are just every as the brain

is able to

try to hack you'll system to perform at that

it can i have a a a really do a an immediate problem in a uh a a where E

it is gonna depend of your uh a now you that you're gonna can see there

but uh this is some ones that would try so two

hard your system

that that's sorry you have to consider that yeah as a bunch of my tell you or

than that that you can be for example the what tell the code they'll detect detector then you would be

able to yeah form more what we code or equal attack

um

here we go that can see there's that's the

at necessary will

have i con turns that that what they mapped to read the

and a a what a map with the same key

you can have all of L scenario where you assume that the just every for example knows the ms that

are embedded in so but a map content T pretty depends of the signal you

and a strong assumption is that uh you assume that to what am is a um uh

that that sorry know everything about

the

but they marking scheme except the secret key

this this is also one the or so as assumption

and what are than the is objective of this uh the adversary re

in all cases than a be to estimate you secret key and uh if it's it's to make what they're

are all copies of sage

okay

and that got that don't actually a lot about what is going to a a a a what the is

that the stories is uh doing to do because it in all case we can that

the uh uh design secure embedding scheme

so uh

here are the assumption that we are are using

we can see they are that the almost is good shown

uh is C is this makes sense to sends to the some limit though is your M of course

um and there's all assumption is the fact that uh as a horse is i I D in the

in there

i and in

plea uh distributed

we assume that all the content uh what them up using the same secret key

and is this means that there is a and that that board W A a that map content only at

that so we have in this frame

the um

i just necessary

uh as only what am i can turn to does not know about them but didn't nice H for example

or was the or here out to design uh secure scheme

uh and also sorts all to play with a to was also a very strong can trends and what that

masking meanings the

and weighting distortions that has to be minimised

and also to maximise the robustness

okay

i to minimize the or uh let's and the robustness so yeah we use um

the rationale already use since a long time in what they're marking

it's to use in it coding and you set side information from the almost

is come from the famous close that's paper but one dove T paper or uh uh

a what that marking

and uh the I Ds and an I D to generate a different coding rage and that we'll code for

the same may say

and

is an baiting in very simple in this case since you have different and dating regions in a to embed

them a said you go to the close

decoding regions and you

uh uh you're able to minimize the distortion okay

so no i will a some up the stream main we got crime i'm and fall secure them baiting

so the first one that i called it distribution splitting

you are have uh

you want to achieve a

a very strong

statistical property

chord is a perfect secrecy uh in steak and the refugee of take was security in what they marking

it's means that the distribution of the content

uh is the original content so P of X and the distribution of the what um not content i it

E E Y given the key a exactly the same

if you are of these properties and so just a re can not do anything about a uh i can

that that that you'll system then because it doesn't know is a coefficient are what all marked or even

uh is the carrier may age

okay

uh

here we assume that we are doing a binary and baiting in means that actually you can't really

you're your uh distribution of what that map the content

P of Y into two distribution

the one uh for of the ms a a zero and the one for the said one

and of course you have to divide

a two in order to normalize the distribution

uh in a bill to achieve this predicting we go now use a partitioning function uh J G

and uh you will see i would show you an example but basically once you have G A a a

uh G

you can compute the

distribution and fall zero according to this partitioning function then you can compute the D

be shown the what are not good content and bidding one

also according

to this partitioning function

you're is an example of the J function so it's

this

lou

uh

that a wise uh function

and so you see that the and function your is divided so it

in in two parts

here i just

are you

the part for one dating zero but you have also as you a part one bit one so the blue

K here

is the distribution

one you want but you want but you

if you add the two distributions of the chrome complain military terry one for dating one you of the now

have

the

exactly the send distributions and the distribution of the also

okay

so it is is the first requirement distributions splitting

now we have to find a map being in of the two

go from

the distribution of

well can which is gauche shown

with a assumption

to is the new distribution when you want for example to embed bet one okay

we can find plenty of different mappings of course

so

C

we can that holds this mapping being uh function T get beat L T

and so the requirement

as a already say so

you have to be to to this

cribbage

so the requirement is that you want to find this because the mapping that will minimize embedding dating distortion

would do we so

so this is important because you have

but you can invent of different mapping but is only one that's

will minimize the average at two T distribution

and here we use optimal once spot in order to do this

so the optimal tons pulse you re give a uh implicit form we'd out of the being in the is

the case of a scan a distribution so in in one D

it's a given by the

mean it

relative distribution function cat be tied at all of the density of

the what that mark content

uh

i

that is applied on the cumulative distribution function of the host content get beat that F four S

this is the that

the mapping from your

you can also derive a is the and bidding being distortion it's right then here

and so you can

once you know the distribution and you can compute the mapping for

he distribution uh you you shows them

uh we're is an example so still a my oldest below

the be shown than the target distribution and one dating is you were

you a a one example of such a mapping here

so this mapping being will and able to minimize at this

the the distortion on average

now

is that everything is done you you you you you you can choose different partitioning function of your distribution you

know either

to uh that form baiting

this is what was called the a once upon that you are what our marking

so you the distribution is very simple if you want to one bad you will all the cool quite if

what are not

a coefficient of an i on the left side of the potion distribution

if you want to embed one

it gonna be on the right side

this is not in at because you have only one a coding region of dictionary region

uh and it was a it comes from a use what

uh no or you can

play with is the distribution and design new distribution here is an example where a or coding and as the

set an uh are and we also the same probability

so the the a here is the same than this one and so long

so i called E

he not you are what that marking be close all

the distance is the

region at the same probability P

in all the two

and as its and probably T P yes

uh you can also see me try a seem that tries this uh D

the and then you are

P bar

that you are what they marking it exactly the same than the previous one but

i perform a tree according to a a or or uh around zero

okay

yeah again i managed to a of the pleading distribution requirement a and and they able to to from and

dating

the last

partitioning function i tried that

is called

the no

that natural are what a marking where a or or then beating uh region as the same way

that is code then does so it's very similar you know to this

can of post test scheme all to you a i am but on you know you have this uh secure

and bidding in this case

okay

so i will now uh compels of defer and to uh what a mapping method the a regarding uh a

tell of eight okay

i want to evaluate knows the last constrains the robustness so i will uh i'm but now assume uh additive

white caution noise uh channel

and here is the compare is and between the that than that you are what that marking and france pop

net you are what that marking needs

just

to evaluate

the benefit of doing

inform it coding

and as you can see so

he's low well well a plot is for

that than that you are what that marking so you can see that you can that shape

uh small of bit error rate

once you can see that in front of coding spatially for lower that uh the value you in a

the you and a the what them map and of the ratio of cool

so

this is so of first benefit

no if i can they so as a different

uh and they so that than that you to what they marking P that you are what that mapping and

P bound that you are what yeah marking

again a again but if they are and what are marked to knows the ratio

you can see is that for example in this case we have a what they map to content which sure

of minus five db

in this case so first the P uh and that you are without marking

gives the best yeah formants below minus five db and that the are you have to use the P bound

that you are what that marking in of the two

uh degrees

you're a bit error rate

if

i i degrees again to what that mark to content issue

i i uh no sorry yeah

um

i wanted also to compare now we is and secure but known to be robust what the marketing scheme so

you have to mail for

to more plots

is that

got more less picky read for uh in bits picked one here

so you can see for example that the improve

a like all for of course low well bit error right for the value and and

and this get us but that's scheme E which fill also

low low well it or right so

oh

is this is a kind of a known in in the secure what they marking it's very out to be

both robust and very secure

this the rubber scheme yeah wow

a more robust and the secure one

and knows of embedding distortion and minus

eleven D V yeah you can see is that

main is the

uh that than that you are what that marking out their forms

the was a a a a uh and dating scheme

so the bit the it much real well for this one

if again i compare our we is this can ask what S T in uh and uh

is the input speech spectrum

i can see is that

the the kind of "'cause" that's in give

even beta a bit error rate

even if

as a at the the value and they to zero you are very you are very close performance as

so

it is also interesting

no i will conclude group michael so

um

it's its importance so to perform distributions coefficients so to split your this

or and you know up to achieve a

secure and baiting

you have also to find a way to match the two distributions so the one of the all sins one

of the what they not go though

in order to do so uh a to do this you can use of to much from sports

it's very it's walks fine in one dimension mention

is the problem is that in a a multi dimensional it's more complicated but you can use all also of

optimization trees and mean know up to minimize the distortion

uh there is a gain of using in from it in also for secure and baiting

and uh and thought and that is so the best partitioning depends of the and dating

used option the but use yeah and also of the noise

so i

we don't have a a a a a magic old up to perform both secure and and and dating

the best pick either a fell to link

what we can do using secure adaptation of the get us but that's you

so they have been applied by around the last week and is using at at at that position of a

uh this yeah that's but that's scheme

and they are or so a more fundamental problems as like i would to compute exactly theoretically is a secure

get by is so

how much information they can convey a while keeping being uh security

and also

another problem is that

if you want to what secure embedding betty a secure uh and baiting uh

with is low bit error rate you have also to use secure uh iraq weighting

yeah well only can code

and this is uh not known right now also we have to find a way to

improve the quality of cell is but also guarantee security it's not easy because if you use uh

they're all calling

encode you gonna add dependence between the

symbols that you're gone and that

and this dependence can be a security uh equal

thank you for your attention

and the time for a couple

that to a kind of nice the

didn't on the first like to set you word and work on a C is just a done the quite

sure what type of security does your

net the dress

and and is you cure in the first place because you kind of

i use you could in first two slice

you never argue why he's

Q also okay

so uh

here are the signal you is that so that very will have only what a market the content and we

try to

it's steam so you will have a bunch a

what a market contents

all more more to be uh what L mark only a set of then will be what they matt

and we assume that all this can this as a cushion and and i a E okay

so the goal of that the we will be two

for example is to the set

oh

seven

that uh

convey

so what am a

so if you want to do this

i i has to use some kind of uh uh security at that it can be for example looking at

the distribution

fine cluster

yeah from independent can it uh and component and that is this would you have

very very different way to perform from uh what uh what am looking at that in the security uh a

lot at the security level

here

we uh i assume that

is there is not possible at that

since

all the coefficient have the same distributions so what it is that the sre

won't be a bore

to know which coefficient

i what alma

and still not quite a what what

what is the goal of the attacker

i'm still not getting a

so what is that the vet analysis that you trying to

event of time

so it's to if you at this you know you a few out of the N coefficient and only you

set of and coefficient

conveys or information

but

is the N coefficients the location of this and clean fish and uh come from us

secret key

so the of the idea so you will be two

uh try to estimate

which coefficient fusion

conveys or information

okay

so

and then then the next

why that's the case why is this in system because it didn't see any secure you know

i only solar boston's in

yeah the security is grounded because you assume

because uh with the em baiting

you have exactly the same distribution before and after of baiting

it's means that

if i want if i plays a but yeah i and and the i am that that's every

i won't be a ball

to locate

which coefficient carries a what that matt so if i want to

to try if i tells a say J you have to add the noise on all the coefficient of my

team

i cannot say act

which one carries a a age

as a a just every

a left so offline

yeah

right

so thank you know match again