Wired article on "Phantom" image hacking of Tesla and MobilEye ADAS

[Soda Popinski]

Link to the article.

It mentions the driver wouldn't know why the car reacted, and there wouldn't be any evidence. I don't completely agree with that. The 0.42 second flash needed to fool a Tesla would be noticeable to the naked eye. If the car has Dash Cam activated, that would also be evidence.

Cruise mentioned a Lidar based system wouldn't be fooled (even though they also use cameras?). I suppose it could check camera data against Lidar data to see if it's really there. If that's the case, I'd expect Tesla's FSD could do the same if it does indeed end up using Pseudo-Lidar. Same with MobilEye's VIDAR (assuming it's used for SLAM and not just generating HD maps).

 

[mikes_fsd]

This is where the 4th dimension of Time is of huge importance.
When you run NN's independent of time, on image stills you are susceptible this hack.
When you are running a NN on video feed and correlating it in time, you effectively can eliminate this to make sure it is present across multiple frames (i.e. time).

 

[heltok]

I think the 4D solution, where you over time paint a birds eye view will be much harder to adverse attack than when you try to fool an image classification by adding a texture patch to the image. I would like to see the attacks and how successful they are...

 

[quadroplexor]

Fixed in FSD rewrite.

 

[heems]

What about if we rent one of those flying gliders with Captain Morgan ad banners being towed. Then we add a stop sign to it. Then fly low over highways. Omg. That’s it. We hacked FSD. It’s not safe!!! Get out of the death trap now while you can!

 

[Soda Popinski]

Fixed in FSD rewrite.

Fine for FSD, but could still be an issue for regular AP folks. Not sure if the re-write will work on HW 2.5.

 

[mikes_fsd]

Fine for FSD, but could still be an issue for regular AP folks. Not sure if the re-write will work on HW 2.5.

Legally, on non-FSD cars, you - as the driver - will always be responsible.
So, keep your sentry mode on and plenty of storage to record all the anomalies.

But, realistically, I see the regular AP (2+) getting a smaller birds eye view build of the NN's that only focuses on the forward facing cameras and specifically target to make the AP features better.
This would be their way of keeping driver assist features ahead of competition.
Every time someone comes close with L2, they will make sure Tesla AP (non-FSD) is able to stay ahead. (my speculation).