Okay.

So welcome to Multimedia Security.

Last week we started to look a little bit more closely into these requirements that

we impose on the quality of a watermark.

So these quality criteria are fidelity, robustness, security, and the payload.

So fidelity means that we would like to have a watermark that is ideally invisible or at

least visually unobtrusive.

Robustness means that it is hard to remove.

And then when you look at the papers you will see like oftentimes it's mixed with security

because this removal can either be accidental so that in all cases is robustness.

But watermark removal can also be due to a targeted attack and then in this case one

could say okay it's a case for security like watermark security.

Or one could say yeah okay so if I like to if I say I'd like to have a robust watermark

that just means that includes robustness and security.

Okay and third is the payload.

So how many how much information can I embed in a watermark which might then be interesting

for example to track specific customers or identify specific items using that watermark.

Alright so I think we stopped somewhere in fidelity is that right or not?

Okay no one knows.

Let's say we stopped in fidelity.

So really quick recap so in order to make sure that a watermark is visually unobtrusive

we said okay we have some embedding scheme for example an additive scheme where we create

a watermark by adding a little bit like some alpha with some alpha weight that watermark

to every pixel.

Or we could also do that multiplicative or so it doesn't really matter.

So we could also say we perform a multiplicative scheme.

That doesn't really matter so in terms of fidelity what is interesting is this alpha

here because it determines how strong the watermark is.

And so the alpha determines the strength of the watermark

and this of course influences the parameters we're interested in and particularly fidelity

and robustness.

So the idea now okay naive approach is okay let's just tune that alpha to our needs but

then in order to improve a little bit over that what we can do is that we define alpha

to be spatially variant and to adapt to the content of the image.

So we can choose a larger alpha in regions where our eye would not notice any differences

and in regions where our perception is very sensitive we tune that alpha down and that

way we can embed in total more into the image at the same like fidelity.

So the idea is create a perceptual mask

to locally adapt alpha so therefore we have an alpha eye instead yeah because it varies

from pixel to pixel and our overall equation then becomes something like okay I mean that's

a no-brainer yi equals xi plus alpha i times wi something like this for the additive case

for example.

Okay and now the question is what is a good perceptual mask and a lot of masks have been

proposed you will at least you will deal with a minimum of two masks when you go through

these papers and all are a little bit heuristic but influenced of course from results from

psychology and let me write down a couple of rules of thumbs how to define a perceptual

mask here how to define a perceptual mask so some rules of thumb

we need need function a function that considers local contrast

and so with the goal of determining regions with low contrast because the eye is more

sensitive there and there we make the watermark weaker in the other regions we make it stronger