 versions using frequency analysis. OK, thanks. Hi. So let me familiarize you with the setting here. So on the left-hand side, we have the sender of this protocol. And on the right-hand side, we have the receiver. And the sender is just some acoustic source. And it's going to transmit some sound waves towards the receiver. So what is the channel here? It's just like the air that we breathe. So as you can imagine, this channel is possibly very noisy. OK, so to see that this noisy channel can be a problem, let's consider a setting that you might be familiar with. So let's assume that this is you. This is the audience. Or probably this guy is you. You just gave an amazing talk at the conference. And then there's Q&A. OK? So of course, the guy or the girl who is asking a question is not going up to the mic. So it's just going to stand up and shout or tell you the question. OK? So it might be that someone is cuffing. The questioner is like in the back. There's some like maybe the architecture is not that good. So it's actually the case that these waves are going to be distorted. OK? And, well, it leaves you, as I speak on the stage, in a not so nice position. OK? So we have actually no clue what the question was about. And this turns out to be a pretty bad problem because generic solutions turn out to be not working all the time. OK? So let's see some examples here. So for example, if the generic answer, this is a very interesting question. Maybe some of you used this already. But what if the question was, have you tried it with obfuscation? Well, another example of a generic answer is, well, we haven't really looked into that. OK? But what if the question was, can you please show us the proof? I mean, unless you work in certain areas of crypto, this is not a good idea. OK. So another example is, let's take this offline. Seems to be always a good idea. Well, except if someone said, well, you should have cited my work. In this case, I can tell you that if you see this sign, these people know where the next hospital is. Actually, they take you to the next hospital for free. No, sorry, not for free. OK, so our observation is that cryptographic questions actually tend to have pretty low entropy, OK? So let's look at one example. Can you please explain the complicated part of the proof again? So actually, the only word that you actually have to get from this is proof, OK? So you're fine. You just explain the whole proof again. Totally fine, OK? So our approach is actually to decode only the crucial words in the question, OK? So I don't want to bug you. I don't have the time. I don't want to bug you with the decoding techniques. It's actually involves going from 2D to 3D. So yeah, it's always a good idea. So our results. If you, I mean, I'm going to just show you like the top words that we managed to decode. If you hear something like, save the nation. Where's the station? You should take some. Then the question was probably about obfuscation, OK? Top two, where's Tad? Who's wet? Question was about the proof step, OK? Top three, East Germany, Patricia Tree, and the Fish Speed. Question was about efficiency, OK? And this is amazing, because already covers like 90% of the question. So I strongly advise you to approach me for the complete list and you should have it at your next talk. Thank you very much.