Human Side-Channels: Behavioral Traces We Leave | Hacker News Reader

Human Side-Channels: Behavioral Traces We Leave

darkreading.com

51 points80mph9 comments

Back when naval communication was radio communication of manually keyed morse code, one major source of naval intelligence was recognizing the "hand" transmitting messages via the timing patterns of key entry. The messages were encrypted and that wasn't broken, but the timing patterns were essentially in the clear and gave up some useful information.

I wouldn't be surprised if you could generate signatures from the timing patterns of text input events in javascript, too.

> I wouldn't be surprised if you could generate signatures from the timing patterns of text input events in javascript, too.

If I remember correctly, some MOOCs make their users type out an honor statement in lieu of a handwritten signature, and record the timing information as you suggested.

The article refers to other behavioural traces in addition to forensic linguistics.

I imagine these refer to aspects such as how and when we login to systems, our patterns of interaction with everyday interfaces, paths we typically take online and so on.

Seems a fascinating area, anyone with other resources on this?

There was a startup a while back making a login auth system based on the gait of your walk. Apparently phones' accelerometers are good enough to throw all the data at some ML stuff and uniquely fingerprint people based on how they walk.

This is also the idea behind Google's click-to-confirm-I'm-not-a-robot captcha system... they track everything from how you move the mouse to how you scroll the page to differentiate bots from humans.

Seems kind of rough....so you receive a ransom email and you do a fuzzy grep of the entire internet to see if forum posts with a similar "linguisitic signature" exist anywhere?

You'd have to narrow down that search significantly to get any results in any kind of reasonable time.

It depends on the corpus. Stylometry (which is the actual name for this, not 'forensic linguistics') can be surprisingly accurate if it has a good selection of text to work with. How hard do you think it would be to crawl public facebook posts, reddit posts, posts to various hacker forums, and build a corpus large enough to connect the dots? I am betting it is a lot easier than most people think.

Next steps will be application of GANs and similar tools to stylometry to either give your posts a linguistic signature of someone else or to just remove enough unique data points to obscure the source.

Tangentially on-topic, but with all the popularity of style transfer nowadays, has writing style transfer been done yet?

http://twin.sci-hub.se/6607/74f935079576886e116256e5869ec69b... Albeit primitive by today's standards, it impressed me at the time.

Funny they say that, considering much tracking their site has....

Reply on news.ycombinator.com