AI researchers claim 93% accuracy in detecting keystrokes over Zoom audio | Hacker News Reader

AI researchers claim 93% accuracy in detecting keystrokes over Zoom audio

arstechnica.com

27 pointsjdmark10 comments

Given the wide range of garbage microphones, position thereof, background noises, audio compression and noise cancellation I have a hard time believing that this attack would survive a real world test. Of course, if you olace a relatively high quality device like the iPhone on the same surface as the keyboard in an otherwise quiet room qnd train your model on exactly that this may be possible, but that's not how most Zoom calls are conducted.

On the other hand, attacks only get better. You can reasonably expect future attacks to account for a larger variety of microphones and environments

As much as that may be true, I'd much prefer for this kind of reporting to include a real-world reproduction instead of the sensationalist 93% reporting.

Or, if you use a desktop, build your own custom keyboard such that there is no hardware benchmark to work off of. Or use a keyboard which has a different key arrangement that alters the timing and press force of keystrokes, such as Dvorak.

Even better if you have a buckling spring or beamspring keyboard, where each individual key can produce a distinctly different sound profile depending on how they are pressed.

Plus, password managers provide a keyboard-free manner of inputting passwords. Why we don’t have 100% usage of high-quality password managers like 1Password or BitWarden is beyond me… it truly is worlds easier than trying to remember passwords, and infinitely more secure if you use thoroughly random passwords that are at least 32 characters (or whatever the max is for the service you are logging into).

Nothing comes even close to bulletproof, but a lot can make things quite difficult at the far end.

One of the favorite features of my custom keyboard is passsword hot keys.

with hot swap keyboard, silent switches on vowels, R, T, most common letters and clacky switches elsewhere could (plausibly) defeat this profiling (or white noise generator)

https://news.ycombinator.com/item?id=37013704

Given that this seems to be based on key placement, I wonder if the model can adapt to international layouts or alternative layouts like Dvorak.

I initially had the same thought as I have a custom layout but realized that once you have identified which key position is being pressed you can then apply standard cryptographic analysis techniques to get the key mapping.

That could be, but depending on the model architecture, it might depend on multi stroke sequences (similar to autocorrect on a phone) rather than identifying each individual keystroke in isolation. If that's the case, the layout may be baked into the model.

Reply on news.ycombinator.com