Music

Alright, so welcome to multimedia security.

Last time we started to talk about image forensics

and about one detection approach to a very straightforward type of manipulation,

so-called copy-move-virtuaries.

So we are talking about copy-move-virtuary detection,

sometimes abbreviated CMFD.

So a copy-move-virtuary, or sometimes called also copy-paste-virtuary,

is a copy and maybe some optional scaling, rotation,

within the same image.

And meanwhile we know the Iran picture, but there's a number of such examples.

Okay, so detecting this type of manipulation is particularly easy because all the information that is required is present in the image.

So the detection approach is to find a large, whatever large means,

area in the image that is very similar to another area.

So it's a search problem, basically.

And in detail, a very generic way of looking at the detection pipeline is in that paper that we actually wrote,

and I've put it online and stood on. Let me make this a little bit larger.

That's too much, no? Yeah.

So the idea is to perform, like to make design decisions for these steps along the algorithm.

So there's a pre-processing step, for example, convert the image to grayscale or split the color channels into three or something like this.

So there's a pre-processing step, and then either we're detecting key points and then we perform a key point matching pretty much in the sense of,

like in a computer vision sense, like computing SIFT features then on these key points and matching them.

Or to just look at image blocks, so overlapping blocks of the image.

So either we use key points or we use a dense grid of blocks and we compute features on top of them.

And then we have features for every little area in the image, region in the image. So the size of such a block is typically really small.

So 16 by 16 pixels, something like this.

Then we do feature extraction, sorry, then after feature extraction we do the matching.

That means we just, for each feature we create a list of similar features.

We can say, for example, we take only the most similar feature or we take a couple more to be a little bit more fault tolerant.

Yeah, and then it turns out, okay, so of course there's one closest feature for each block or each key point in the image.

But then we need another filtering step to say, okay, we are only interested in groups of blocks that match with other groups of blocks,

that we have larger areas. And that's sort of a noise suppression step.

And this is actually really hard to describe formally.

I wrote it once in a paper totally wrong and luckily the reviewers did not look at this part so much.

And then only, so they accepted the paper. And then for the final version, like there's another thing like, yeah, okay, can you please put some corrections in and so on.

And then I realized that I totally messed up describing this filtering step and then we fixed it.

Okay, so it should be okay now in the published version.

But it essentially says if I have a pair of matched blocks and then if I look at another block that is close to the first block in my pair,

then and I look at its partner, then this partner should be close to that partner or so.

So that was what I was trying to say.

But it became much more complicated than just saying, okay, I have a group here and I have a group there and all of them have to match each other.

Yeah, okay, good. And then you can optionally do some post-processing, for example, apply a morphological operator to close the mask or so to make it a little bit nicer.

So that is a very generic approach that essentially all algorithms use.

And the question is only, okay, what design decision do I do specifically in each of these steps?

So let's sort of copy that figure really quickly.

So we have a pre-processing step.

We either look at key points or at blocks.

So block tiling and then we have the feature extraction

matching of the features and the filtering of the matches.

And finally, it's not so important, some post-processing.