Hello and welcome to the Thursday, September 28, 2023 edition of the Sansonet Storm Center's Stormcast. My name is Johannes Ulrich and today I'm recording from Jacksonville, Florida.

But we got an interesting new side channel attack. In this case, it's affecting GPUs, not CPUs. Of course, we had a bunch of

different side channel attacks over the last few years in CPUs. Well, GPUs, overall, shouldn't

really be all that different. And so no real big surprise that it affects GPUs, but the kind of vulnerability here is

sort of unique to GPUs.

GPUs are often compressing traffic.

Of course, video traffic is known to compress fairly well.

And as so often, the ratio at which content compresses allows you to deduct

some properties of the content itself. What makes this paper even sort of more amazing is that

they actually managed to implement this attack in a web browser. Web browsers are trying really hard

to separate data loaded from different web pages or different origins. As a result, if you have

two windows open, well, JavaScript in one of those windows cannot access the other window if these two pages were loaded from different origins.

What they demonstrated here in the paper is that they were able to load a page.

They used Wikipedia as an example into an eye frame, and then the page that contained the eye frame was able to

deduct content of the page inside the eye frame, which, well, shouldn't be possible in a sort of

proof of concept. They were able to read the username. Now, they picked the username here because

it wasn't a very specific location of the page. This attack is like many side channel attacks,

not very fast. Speed and accuracy depends a lot on the screen resolution as well as on the

CPU and GPU being used here. The fastest they had was two pixels per second, but with a 99.6% accuracy. Some of the slower ones are going down to like 0.2 pixels per second. So basically it takes you 5 seconds to deduct the content of a single pixel.

Still for a proof of concept, that's certainly not bad and certainly useful to, for example, read content,

like in this case usernames or other confidential data from another page that the script running on the attackers' web page should not have access to.

I will, of course, link to the paper in the show notes.

...