Interesting Research on Subtle Vulnerabilities of Autopilot Lane Detection

[berkeley_ecar]

A recent article in IEEE Spectrum discusses research into vulnerabilities of Autopilot lane detection. Given that the article is dated 1 April 2019, it may be a rather sophisticated joke, but it reads as if in dead earnest (pardon the pun). Perhaps some of our house exports on neural network technology want to chime in on this?

 

[berkeley_ecar]

A link to this report was reissued today in email to IEEE members, dated 10 April -- I'm thinking this is likely a genuine study. This work, out of Tencent, demonstrates a subtle vulnerability in which three small white spots, properly placed in an intersection, can cause a vehicle on Autopilot to swerve into the wrong lane. This is not an immediate practical worry, but does show that the complex "black boxes" that we call neural networks can be fooled into doing the wrong thing. I'm sure the folks at Tesla are looking at this work carefully.

 

[jdjeff88]

This was widely reported more than a week ago and I believe Senor Musk has tweeted about it as well.

 

[MarcusMaximus]

A link to this report was reissued today in email to IEEE members, dated 10 April -- I'm thinking this is likely a genuine study. This work, out of Tencent, demonstrates a subtle vulnerability in which three small white spots, properly placed in an intersection, can cause a vehicle on Autopilot to swerve into the wrong lane. This is not an immediate practical worry, but does show that the complex "black boxes" that we call neural networks can be fooled into doing the wrong thing. I'm sure the folks at Tesla are looking at this work carefully.

One note on this is that similar dots would likely fool actual human drivers as well. When you’re literally changing the direction of dotted lane lines(which this does), bad things are likely to happen.