With 90 minutes of data, they could pick out one driver out of a sample size of 15. How meaningful is that, really?
Honestly it seems like everyone reading these kinds of papers/articles does a ctrl-f "sample size" before anything else. Bit cynical.
It's been easy to do face identification for a very small population for a while. But scaling that to thousands of people has proven quite difficult.
https://www.theguardian.com/technology/2016/may/17/findface-...
and the HN discussion
Modern cars will already learn about driving characteristics so they can better manage the car to match, but I don't know if they try to match individuals.
My truck has three "memory settings". I can adjust everything the way I want it -- different seat positions, mirrors, etc. -- and save those in "A". My girlfriend can then get everything the way she wants it (she's much shorter and smaller than me) and save her settings in "B". When I get in to drive after she has been driving, I just hit the "A" button to have everything adjusted back to exactly how I want it.
It's very handy when you have multiple people who drive a single vehicle.
I rented a new Chrysler 200 a few months ago and just when I thought I had the throttle figured out to where I could trick the transmission into not upshifting when I wanted it not to, upshifting when I did want it to, etc... it stopped working. It was a bizarre experience, like some sort of arms race or adaptive-difficulty video game between it and me, with my goal to actually control the RPM and its goal being to prevent RPM from getting out of its preferred small range no matter how increasingly erratic-seeming my throttle foot behavior became.
As a result, I am unable to develop any intuition about how the engine will actually behave in response to my inputs, and in a split-second dangerous situation will be unable to do anything useful.
It's advertised as "helping fuel economy" but all it really is is "adaptively learning to ignore the operator's inputs" which might be nice for the first few thousand miles a brand new driver drives, but seems outright dangerous afterwards.
I really don't know how people drive, let alone enjoy, cars with automatics these days. Paddle shift with dual-clutch autos might be loads better, but they don't generally have those on cheap rentals, so I haven't had much time to play with one.
This has implications for "which family member is it?" or "you claimed you lent it to your friend, is that really true?" etc.
So, the claim is 90% accuracy in determining which one of the fifteen drivers they are looking at. So, they were correct on 13.5 drivers.
Um, what?
I don't see where you're coming up with only 15 minutes total driving. That isn't implied anywhere in The Fine Article. But even leaving that detail aside, it would actually be more impressive than the reported finding if they were correctly identifying people that well after only 15 minutes total!
If each driver is run in twenty 15-minute (or longer) trials, you have 300 total data points to analyze. If the car was correct in its fingerprinting 270 times, then it was correct 90% of the time.
No, it could have been inconclusive for the other 10% of the trips, simply stating, "Insufficient data for meaningful answer". So 90% of the time it said, "Oh, this must be DrScump" and the other 10% of the time it said, "I don't know who this is yet, let me keep analyzing the braking patterns"
EDIT: Which is different than the way I stated it in my initial comment. Your reply helped me realize that.