Skip to content
AI's Whispering Secrets: Typing Detection from Audio Raises Privacy Concerns episode artwork

Jan 21, 2024 14 min

AI's Whispering Secrets: Typing Detection from Audio Raises Privacy Concerns

from The Daily AI

In this episode, we discuss the privacy concerns surrounding AI's newfound ability to detect typed text from audio, whispering secrets that were once thought secure. Tune in for insights into the balance between technological advancements and individual privacy rights. Invest in AI Box: https://Republic.com/ai-box Get on the AI Box Waitlist: ⁠⁠https://AIBox.ai/⁠⁠ AI Facebook Community Learn more about AI in Music Learn more about AI Models

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

AI's Whispering Secrets: Typing Detection from Audio Raises Privacy Concerns

0:00 14:25
of matches

Transcript Auto-generated, may contain errors

Today on the podcast, we're going to be talking about some alarming new studies saying that humans cannot detect deep fake audio. We're going to be talking about AI that is trained to listen to the sound of you typing and be able to detect what you are typing. And then we're going to be talking about a new patent that Apple just filed to be able to pretty much receive bio data from your AirPods. So without further ado, let's jump into these three stories.

So I mean, essentially, I think what's really interesting here is with AI, there's a lot of incredible advancements, but at the same time, in my opinion, there's a lot of big dangers and there's a lot of ways that AI is, you know, essentially being used to get rid of a lot of our privacy. So one of the big things and the big headline articles that just recently came out is that according to Bleeping Computer, there's a collaboration between various UK-based academic institutions, and they essentially use AI and they use a microphone to predict what you're typing, right? So they could just take the audio track of you typing something. This is kind of cool because it would make, like, well, this isn't cool, I actually think this is alarming, but essentially they can use this by listening to the sound.

They can feed the sound of you typing something and then say what you typed to the computer or into the AI model. They can train the AI model off of that, right? That kind of makes perfect sense. They put the audio file and the transcript side by side and they can train off of that.

So I think this thing got really good. Essentially, use AI, use a microphone and it can predict what you're typing. So imagine the implications, right? Everything from your passpodes, your personal chats, your corporate confidential information, all of that could be exposed to an unauthorized audience just by hackers or anyone getting access to the audio.

And if you think about it, this is less difficult than, you know, you might think for, like, you know, trade secrets or something. All someone would have to do, they don't actually have to get access to your computer or your phone. They just have to be able to have some sort of recording device in your work environment, for example, to be able to get access or listen to your personal work information. You can imagine where, let's say there's a notable tech company, Google, and some competitor or, you know, let's say a Chinese counterpart firm that would like some sort of trade secrets.

They go and buy the janitorial company that works at Google's headquarters and they just have the janitor drop off and just leave a specific device in, you know, maybe an executive's office. They say, hey, this is an air purifier. We should just, like, leave this in there overnight or for the next week. Then collect it and bring it back to us in a week.

Bada bing, bada boom, no one thinks anything about it. They take that air purifier, which, in fact, maybe was a real air purifier, but just got a bug inside of it. Or they just stick it in the vent, whatever. You know, you can imagine.

It sounds like a spy movie. But, like, this is really not that far-fetched. The intelligence services, I'm sure, have been able to do this for a long time, right? Every time a new technology or breakthrough from academia comes down the pipe like this, that, in my opinion, gets rid of a little bit of your privacy.

I'm like, well, you know, FBI, CIA, probably been able to do this for a while. But it's just, but so anyways, I think it's great that people know it's now possible, but it is, man, what an alarming concept that literally you typing, like your password's nothing, everything can be detected now. They can literally just read the transcripts of you typing. I wonder if there's, I wonder what ways people try to have to get around this.

They kind of type funny or, anyways, I don't think it's going to work. So it's interesting because there are other options, right? You people do, I mean, really, there's voice to text or typing. That's how you get stuff in.

You can type stuff on your phone, turn the clicks off on your phone. It's going to be interesting. I mean, yeah, anyways, so that's a really interesting piece of information that just came out. And essentially, the researchers carried out their experiment because essentially what it's doing is the distinction between the subtle sound differences that keys make and also the patterns in which you type words, and then it is able to detect them.

But also, it's literally from the subtle sounds that different keys make. That's one of the big variables here. So at first, I was like, well, if you type different, like instead of just typing at your normal speed, you try to like type slow and fast and slow and fast. No, I think it's actually able to literally detect down to the key, which one you're pressing, and it's kind of like a code where once it cracks the code, it knows everything you're typing because it just knows every key stroke.

So very, very crazy. In any case, researchers carried out their experiment on MacBook Pro, and they caught audio snippets from a proximate iPhone. And then here's the twist. Through a Zoom call, they were also able to do that.

So that shifts the, you know, onus of security, not just to you, but on every individual present on that call, right? So you can, they're able to do it from, sure, an iPhone is in the same room as you recording it, but if you're on a Zoom call chatting and you're typing something, like essentially the transcripts, like if someone records a Zoom call, they can literally use that recorded Zoom call to see what you were typing in while you were on the Zoom call, right? You're on Zoom call with your boss, they're talking about an important business thing you aren't involved with, so you get sidetracked, you go to YouTube, you type in something, if you're looking at the score for a sports team, they could run that audio for the call through an AI and see what everyone had searched during the call. Man, that'd be absolutely crazy.

And if you think about it, this isn't just like someone has to manually do this. I'm sure there's all sorts of like enterprise tracking software, especially when everyone's working remote. I'm sure there will be enterprises, there will be software, there will be companies that essentially have, you know, enterprise monitoring where it's like, if people are doing stuff they're not supposed to, people get notified and it'll all just be through like Zoom calls and other stuff. I'm sure like most enterprise, they're tracking a laptop and everything on anyways, but yeah, it's just crazy.

It's an interesting concept that everyone should definitely be aware of that it's now possible. So segueing into the other study I mentioned earlier, I think what's really interesting, there's a study out of the University College of London and it essentially said that humans are not very good at distinguishing real speech from deep faked content. So across English and Mandarin, participants could accurately identify a synthetic, you know, AI generated deep fake voice 73% of the time, which means that almost one in, you know, over one in four instances, humans were not able to identify the deep faked content, which, you know, I know a lot of people be like, hey, that's not bad, right? Like, you know, scammers and all sorts of people, you know, deep faking content on the internet, whatever, you know, if we could detect it 75% of the time or almost 75% of the time, right, that's pretty good.

But I think what's really important here is we know that the quality is only going to improve and this is an area that is constantly evolving. And so if we can, if we're already, if we've already lost 25% of, you know, our detection and being able to understand this, I think it's a slippery slope to it's 90% undetectable. It sounds exactly like a person, the person we know, the person we're used to. And there's a lot of implications, but I think, you know, one of the really daunting kind of takeaways here is that deep fake technology in this study wasn't even the latest and greatest.

So as AI speech synthesis continues to refine, I think our human ears are going to face an even steeper challenge in distinguishing fake from real. So Kimberly Mai, who's the lead author of that particular study, she highlighted the need to essentially advance automated deep fake detection systems, especially since deep fake audio gives away even fewer signs than its video counterparts, right? Like if this is a deep fake video, it's a little bit easier for you to see a problem between the audio and the video. But when it's just audio, I think humans really struggle a lot of times.

So interestingly, the study found that participants' intuition played a significant role. Often terms, they used words like naturalness or robotic, and that was what essentially drove their decision-making into picking the human voice or the non-human voice. So when pulled together, collective human judgments termed as, you know, quote, crowd performance essentially matched the presuppiciency of top-tier automated detectors, which is really interesting, right? Some of these top-tier automated detectors can get this amount of time, but when you have a few people listening, a few people would be able to detect it most times.

So I think in today's really rapidly advancing landscape, this is something we definitely need to think about and address. I would assume that this is especially prevalent for elderly folks or people that aren't, you know, as exposed to the advancements that are happening in AI and in technology. And so I think it's just important that we, you know, make them aware. The last thing I wanted to bring up on the podcast today is the fact that Apple has just filed a patent with the US Patent and Trademark Office, and they were granted essentially the ability to create AirPods that will be able to read biosignals emitted from your body.

So this is going to be done by using the scattering of tiny electrodes through the earbud, which will facilitate biosignal tracking. So in addition to this, AirPods are also going to contain active and reference electrodes that will be customizable based on an individual's ear. So apparently in the, in a report on BBOM, it said, quote, these electrodes will be present in both wired and wireless earphones from Apple. The primary use of these electrodes will be in the form of electrophilography, EEG.

I may pronounce that wrong, but it's EEG. And the patent also reveals that the lower most section of the stem can contain an electrode that the user can easily press to trigger the start of an EEG session and other related measurement activities like electromyography, electrochorography, electrocardiograms, galvanic skin response, and blood volume pulse measurements. This is absolutely insane in my opinion. This actually goes back to something that I've been talking about for quite a while since the beginning of the podcast, which is that eventually they're going to stick fMRI scanners into VR headsets to be able to literally read your mind because with an fMRI scanner and an AI model, you can literally read someone's mind, the exact thoughts that they're thinking, the exact words that they're reading.

If you just train a model based off of someone's, like they train a model off of someone's brain waves while they were reading something, they fed the script in on one side and the brain wave scan on the other side. And that model can now literally detect somehow pulling full sentences word for word straight out of your brain as you think them. They can pull video and images as you imagine them out of your brain and use something like stable fusion to generate it at a fairly highly accurate level. And so in any case, in my opinion, it's going to be a safety feature, right?

Like, oh, we need an fMRI scanner in there to scan your brain to make sure, you know, it's essentially like the way your Apple watch tracks your heart and stuff and it's, you know, saves people from having a heart attack. That sounds great, right? So an fMRI scanner will save you from all the things you need an fMRI scanner to scan your brain for and to make sure that you're healthy, right? But also if someone hacks into it or if, you know, let's say the government gets a warrant and requests access to your fMRI scans and then literally can read your mind, to me, that sounds like a pretty big breach of trust.

And I don't think it's far off. I definitely think that's going to be something that will be added to VR headsets in the future because those go right over the top of your head. And that's kind of the perfect way to implement a technology like that. Now, that was kind of my original fear, but where I think this probably goes is it gets to a point where your AirPods and your earbuds will probably be able to read your mind at some point.

And it sounds like so conspiracy, so tinfoil hat, so crazy, but I almost can guarantee in my lifetime and probably sooner than later, your AirPods will be able to be an fMRI scanner that can read your mind. And there may also be other technology that is getting introduced. There was a lot of tech that was going to, the Apple's already working on integrating. And then I just think it's crazy that they're already doing it in these AirPods.

You know, it'll be even easier for like an fMRI scanner, for example, it'll be a lot easier with over-ear headphones, where it's actually like has a band that goes over the top of your head. That'll be a lot easier for an fMRI scanner, something like that. These AirPods would definitely be a little bit of an engineering feat. I think it's going to get there.

I think they're going to be able to do it. And I think it's going to be absolutely insane. So that is all for today on the podcast. This is definitely areas that we're going to be following from audio to visual to all sorts of areas in essentially hacking and tricking the human mind with AI.

There's a lot to be aware of. So that's what my goal is today is to bring you awareness of what is currently happening so you know and can take necessary precautions. I mean, it's crazy. There's a lot of implications.

I would love to hear your thoughts on this and how you think this is some implications I may have missed. Feel free to share those on our Facebook group. But otherwise, have a great day. Thank you for listening to today's podcast episode, breaking down how AI is impacting your industry.

Today's episode is sponsored by AI Box, a no-code AI app builder and marketplace, which just launched a crowdfunding campaign. If you are interested in investing in a new AI startup, you can go to republic.com slash AI-box. I'll leave a link in the show notes as well to learn more. The minimum investment is $150 and the maximum investment is $100,000.

Until next time, have a fantastic day.

Frequently asked questions

How long is this episode of The Daily AI?

This episode is 14 minutes long.

When was this The Daily AI episode published?

This episode was published on January 21, 2024.

Is there a transcript available for this episode?

Yes, a full transcript is available for this episode. You can read the complete transcript on the episode page.

Can I download this The Daily AI episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.