New acoustic attack steals data from keystrokes with 95% accuracy

L4sBot@lemmy.worldmod to Technology@lemmy.world – 441 points –
bleepingcomputer.com

New acoustic attack steals data from keystrokes with 95% accuracy::A team of researchers from British universities has trained a deep learning model that can steal data from keyboard keystrokes recorded using a microphone with an accuracy of 95%.

70

Iโ€™ll believe it when it actually happens. Until then you canโ€™t convince me that an algorithm can tell what letter was typed from hearing the action through a microphone.

This sounds like absolute bullshit to me.

The part that gets me is that the ONLY reason this works is because they first have to use a keylogger to capture the keystrokes of the target, then use that as an input to train the algorithm. If you switch out the target with someone else it no longer works.

This process starts with using a keylogger. The fuck you need โ€œaiโ€ for if you have a keylogger?!? Lol.

That's pretty much what the article says. The model needs to be trained on the target keyboard first, so you won't just have people hacking you through a random zoom call

And if you have the access to train such a model, slipping a keylogger onto the machine would be so much easier

Hmmm not totally. A bad actor could record the keyboard and then figure out a way to get it installed. Either through a logistics attack (not everyone maintains a secure supply chain), or an insider threat installing it. Everyone's trained not to allow thumb drives and the like. But a 100% completely unaltered bog standard keyboard brought into a building is probably easier, and for sure less suspicious if you get caught.

Sure you might say, "but if you have an insider you've already lost" to which I say, your insider is at risk if they do certain things. But once this keyboard is installed, their own detection risk is less.

Now the question is, how far away can the mic be? Because that's gonna be suspicious AF getting that installed. BUT!!! this is still a great way to break the air gap.

A bad actor could record the keyboard and then figure out a way to get it installed

The room is important to the training of the model as well. So even if you know the make and model of the keyboard, the exact acoustic environment it is in will still require training data.

Also if you can install a keyboard of your choosing, you can just put the keylogger inside the keyboard. If you're actually getting your own peripherals installed on your target machine, training a model to acoustically compromise your target is the most difficult option available to you.

good point about the room.

as for an installed keylogger, there are organizations that will inspect for that and catch it. My point is this is a way to get an actually unmolested USB device into play.

But I hear you, this isn't likely an ideal option right now, but it is an option for maybe some niche case. And these are early days, put enough funding behind it and it might become more viable. Or not. Mostly I'm just offering the thought that there ARE use cases if someone puts even a moment's creative thought into trade craft and the problems it might solve like breaking the air gap, emplacement, avoiding detection, and data exfil. Each of those are problems to be solved at various levels of difficulty depending on the exact target.

I think you might have misunderstood the article. In one case they used the sound input from a Zoom meeting and as a reference they used the chat messenges from set zoom meetings. No keyloggers required.

I haven't read the paper yet, but the article doesn't go into detail about possible flaws. Like, how would the software differentiate between double assigned symbols on the numpad and the main rows? Does it use spell check to predict words that are not 100% conclusive? What about external keyboards? What if the distance to the microphone changes? What about backspace? People make a lot of mistakes while typing. How would the program determine if something was deleted if it doesn't show up in the text? Etc.

I have no doubt that under lab conditions a recognition rate of 93% is realistic, but I doubt that this is applicable in the real world. Noboby sits in a video conference quietly typing away at their keyboard. A single uttered word can throw of your whole training data. Most importantly, all video or audio call apps or programs have an activation threshold for the microphone enabled by default to save on bandwith. Typing is mostly below that threshold. Any other means of collecting the data will require you to have access to the device to a point where installing a keylogger is easier.

It sounds like it would have to be a very targeted attack. Like if the CIA is after you this might be a concern.

Yes, exactly.

Actually I just saw this: Zoom terms of use updated to allow AI training on user-generated data, no opt-out

Maybe if zoom is systematically collecting data on all users they would be able to build a reasonable model. Then it could be leaked or shared.

What do you think?

Good question. Since Zoom is mainly a buisness tool and a lot if high profile companies rely on it - if there's even the suspicion that zoom uses collected data to steal passwords or company secrets, they will bring the hammer down in the most gruesome class action lawsuit. Companies pay good money for the buisness license and Zoom will certainly not bite the hand that feeds them.
However, this might not apply to private Zoom users. And I'm certain that Zoom does some shady stuff behind the scenes with the data they collect on private individuals beyond simply "improving our services".

I'm skeptical too, it sounds very hard to do with the sound alone, but lets assume that part works.

The keylogger part could be done with a malicious website that activates the microphone and asks the user to input whatever. The site would know what you typed and how it sounded. Then that information could be used against you even when you are not in the malicious website.

Hard to do, but with a very standard keyboard like a Mac keyboard the resonance signatures should be slightly different based on location on the board, take into account pattern recognition, relative pause length between keystrokes, and perhaps some forced training ( ie. Get them to type know words like a name and address to feed algorithm) I think it's potentially possible.

it doesn't need a keylogger. Just needs a Videocall meeting, a Discord call meanwhile you type to a public call, a recording of you on youtube streaming and demoing something.. etc.

Well to train ai you need to known what the correct answer is.

It's bad now, but where we're at with AI... It's like complaining that MS paint in 1992 couldn't make photorealistic fake images. This will only get better, never worse. Improvements will come quickly.

Sounds like a fantastic way to target a streamer, but it's otherwise very limited.

Is gonna sound crazy, but I think you can skip the keylogger step!

You could make a "keystroke-sound-language-model" (so like a language model that combines various modalities, e.g, flamingo), then train that with self-supervised learning to match "audio" with "text", and have a system where:

  • You listen to your target for a day or so, let's say, 1000 words typed in ๐Ÿคท๐Ÿปโ€โ™‚๏ธ
  • Then the model could do something akin to anchor tokens in language-to-language translation, except in this case it would be more like fixing on easy words such as "the" to give away part of the sound-to-key map. Then keep running this mapping more parts of the keyboard
  • Eventually you try to extract passwords from your recordings and maybe bingo

I think it's very narrow to think that, just because this research case requires a keylogger, these systems couldn't evolve other time to combine other techniques

It looks like they only tested one keyboard from a MacBook. I'd be curious if other keyboard styles are as susceptible to the attack. It also doesn't say how many people's typing that they listened to. I know mine changes depending on my mood or excitement about something, I'm sure that would affect it.

Tangentially related: Did you know, that itโ€˜s technically also possible to reconstruct sound via smartphone accelerometers and thereโ€˜s no restrictions on which apps can use it. Have fun with this info (:

Reconstruct sound using smartphone accelerators? What do you mean? That accelerometers can act as speakers and produce sound? Or they can act as microphones and record sound as numerical data of vibrations etc? Can you point me to any articles or sources?

SpyApp is spying in background

User thinks "why is battery draining so fast?"

Opens battery setting

Oh, this app shouldnt work right now

Restricts SpyApp's battery permissions

are you saying that a cellphone accelerometer can be used as a microphone? That sounds... interesting. Do you have a source?

I am not the person you are replying to, but if the accelerometers are sensible enough, the vibration of the voice will be picked up by the accelerometer.

Since the sound we make when talking are periodical, it can probably easier to track that periodicity and reconstruct the sound from there.

It's all my (un)educated guess.

Iirc on newer versions of Android there are restrictions on polling rate of sensor data

laughs in custom multi-layer orthogonal layout with one-of-a-kind enclosure & artisan keycaps

Only plebs type. I write all of my content in machine code with a custom compiler to translate it to QWERTY.

NSA/CIA/DEA/Interpol/FBI still trying to decode my shitposts to this day

You have to train it on per device + per room basis and you don't give everything access to your microphones

I was just thinking, streamers might have to be careful actually โ€” you can often both see and hear when they're typing, so if you correlated the two you could train a key audio โ†’ key press mapping model. And then if they type a password for something, even if it's off-screen from their stream, the audio might clue you in on what they're typing.

Iโ€™m sweating. I use blue switches. Help.

Never knew my mutant blue switch keeb would come in handy one day. I've lubed the blue switches and added foam and tapes so now it sounds like a clicky-thocky blue-brown switches keeb.

Assuming that this does not only work on English words, this is actually really terrifying.

I have to assume it could be modified to work on any language. You just have to know the keyboard layout for the language in question do you know what to listen for. Languages with a lot of accents like French maybe could be slightly more complicated but I seriously doubt that it couldn't be done. I'm honestly not sure how the keyboard is set up for something like Chinese with so very many characters but again if this can be done, that can be done with some dedication and know how.

There are several different ways of inputting Chinese, but generally they all map 2~6 keystrokes to one or multiple Chinese characters, and then the user chooses one. I'd imagine it wouldn't be much harder.

Does the research presume the use of a qwerty keyboard?

It uses the sounds it records and compares again the messages you send. So in theory it's layout agnostic.

Phreaking for the modern era.

I find this article kinda mid bc No link to og paper Article doesn't specify what kinds of keystrokes were being detected (so title seems kind of clickbait)

  • probably not all kinds of keyboards if they only trained model on macbooks? Also no mention of kind of data used to demonstrate 95% accuracy

When your ADHD fidgeting and a mic attached to your head become a super power. No one can read my keystrokes!

Will a faraday bag help with a phone? Seeing how it blocks connections. You can unplug desktop mics.

A very widespread implication of this is if you are on a call with a bad actor and are on speaker phone, and you enter your password while talking to them, they could potentially get that password or other sensitive information that you typed.

Assuming it really is that accurate, a real-world attack could go something like this. Call someone and social engineer them in a way that causes them to type their login credentials, payment information, whatever, into the proper place for them. They will likely to this without a second thought because "well, I'm signing into the actual place that uses those credentials and not a link someone sent me so it's all good! I even typed in the address myself so I'm sure there's no URL trickery!" And then attempt to extract what they typed. Lots of people, especially when taking calls or voice conference meetings or whatever from their desk, prefer to not hold their phone to their ear of use a headset mic and instead just use their normal laptop mic or an desktop external one. And, most people stop talking when they're focused on typing which makes it even easier. Hell if you manage to reach, say, the IT server department of a major company and play your cards right, you might even be able to catch them entering a root password for a system that's remotely accessible.

From the article:

The researchers gathered training data by pressing 36 keys on a modern MacBook Pro 25 times each and recording the sound produced by each press.

In their experiments, the researchers used the same laptop, whose keyboard has been used in all Apple laptops for the past two years, an iPhone 13 mini placed 17cm away from the target, and Zoom.

Now they should do this under real usage and see if they get anywhere close to 95% accuracy. Phones are usually in pockets, people listen to music, not everyone has a MacBook.

I think it will be difficult for the average person to use this attack effectively, but I think this will become some sort of government spy thing for sure.

I'm just going to play a keyboard ASMR video while I type. Problem solved.

That "95%" has about as much credibility and extremely specific test conditions as MPG for cars

I think i saw this in a movie once. Does anyone remember what movie it was?

Can't remember if the one I saw was movie or TV but I want to say the plot device in question (from the instance I remembered) was a small acoustic bug under the caps lock key. At the time I thought that was too far fetched to be possible. We live in a world....

Can we, one day, have a research, a project where DL, AI, LLM (w' the f*ck you call it) solving real and useful problems?

I swear, these techs are boring as f .