Hacker News Reader

Comment by duxup | Hacker News Reader

Is it AI or just ... recognizing a pattern?

How much data could it have to look at in the time that someone "snatches" a phone?

The article continues with:

> If a common motion associated with theft is detected, your phone screen quickly locks – which helps keep thieves from easily accessing your data.

So it's probably a machine learning model that was trained on motion data of snatches, but it's likely not AI in the sense of LLMs.

But I wonder how many false positives this could yield. For example you are in a hurry and you snatch your phone from a table. How precicesly can this model decide with just motion data, if this was theft or not.

I'm trying to decide if parents with small children will either love or hate this feature.

Google is really going to crush those little toddlers dreams of finally getting their hands on the phone :(

Would be interesting to know the difference between a snatch and me rushing out the door...

Do you often interact with your phone via the screen while rushing out the door?

Yes?

Yes. Often the map or messaging app.

Probably the acceleration vector. If the phone is rapidly moved a meter away from you, either it's being snatched or it's being thrown.

Edit: To clarify, I was thinking of horizontally, in the direction that corresponds to the top of the screen, as if you were bent over using the phone--probably holding the bottom-of-screen--and then someone grabbed the top-of-screen to pull it away.

> If the phone is rapidly moved a meter away from you, either it's being snatched or it's being thrown.

Good heuristics. Also that must not be a mainly downward rapid movement, which probably only means you just dropped your phone.

I think a lot of the false-positive cases where the screen gets locked are acceptable in context.

I mean, most people dropping their phone will be too glad/devastated that the device did/didn't escape harm to bother being annoyed that they have to unlock the screen again.

A drop registers as no acceleration while in freefall, and then a sharp spike when it hits the ground. This was counter-intuitive to me when I first figured out how to display my phone's accelerometer readouts, but makes sense.

I mean, worse case scenario, your phone just locks (I assume to the lockscreen, where you have to re-enter your pin). It doesn't seem like such a big problem?

It's a classifier, like ML has done for many years.

There's a saying that when something becomes mainstream it is no longer considered AI. Fun to see that being reversed.

Personally, I would take the false positives. Way too much of my life is locked into securing this fragile black rectangle. Unlocked phone has access to basically everything. I personally do not do any finances on my phone, but all of the MFA works through it.

If I snatch a phone from the table (probably already locked?) or drop it, I will suck up the additional login.

I have long thought about the utility of a little locking-beacon. If phone suddenly gets out of range, should auto lock. If only Bluetooth were not so unreliable.

My Apple phone won’t let me do some sensitive things if I am in an unusual location. It’s a default setting.

If my phone gets nabbed, a motivated thief would do nefarious actions the minute they get out of site. So presumably just a few blocks away from a usual location.

Worth noting Android (or at least Pixel) does have a feature like this, but it actually does the opposite: while a Bluetooth device is connected it keeps the phone unlocked. It would be way more useful in the reverse: that if a Bluetooth device disconnects, it should lock.

These are two different things, since I do not want my phone to have no lock screen just because my headphones are sitting near it, but if it is unlocked and suddenly my headphones disappear, that would be a useful precaution, even if it doesn't eliminate the risk on its own.

You can do something like this with an automation in the Shortcuts app on iOS.

It's called Android Smart Lock and has been a thing since Android 5 (Lollipop). It also works (worked? Does it still exist) with 'trusted places' (GPS/WiFi), Voice/Face (before face id was a thing) and a mode were it kept the phone unlocked as long as you kept it on your body.

I remember that I used the last option many years ago, as that was really convenient and worked very well. Basically as long as the phone was in your hand or pocket it kept unlocked, but as soon as you laid it on a table it got locked.

But now that fingerprint unlock is a thing, I don't even mind unlocking my phone as it is one fluid motion and happens unconsciously.

I would expect this to not make a difference if your phone was already locked. But I guess Google could only lock the device if it was upright before being grabbed.

How on earth do you even get training data for this? Recorded phone sensor outputs that are known for certain to be the result of validated, confirmed theft events can't possibly be that common. Are they paying people in Bangledesh a few bucks to be randomly assigned to group that either get robbed or tripped in the hope they throw the phone and labeling sensor data accordingly. When this type of motion recognition was first developed, they had labs and recorded people walking, doing jumping jacks, sitting and then standing, whatever, and labeled the patterns appropriately because they knew what was happening because it was happening in a lab.

There are free and open source apps for Android that automatically lock the device when the accelerometer detects rapid acceleration, which is a simple detection method. For example, Private Lock is on F-Droid:

- Private Lock (source): https://github.com/wesaphzt/privatelock

- Private Lock (F-Droid): https://f-droid.org/en/packages/com.wesaphzt.privatelock/

Very interesting.

Gotta admit first thing I would do is stage a theft scenario to see how it works.

Linear regressions are machine learning.

AFAICT all machine learning models right now are just pattern matching.

AI is another word for training-based computerized pattern recognition.

Reply on news.ycombinator.com