An interesting anonymization - useless for instant traffic updates, though - would be to fuzz the location based on vehicle speed. When you're moving quickly, very precise data would be shared, but when you're stopped, uncertainty would be increased - that way, where you're parked every night would be much less precise than the highway you take to work.

This wouldn't do much good in more rural areas, of course - you could probably zero in on exactly where my parents park their car with a month's worth of data even if you add a mile-wide fuzz to every parked check-in.

Need some sort of deterministic bias based on user ID? Each user offset 1*Math.rand(hashseed) miles from location at randVector(hashseed)

Why does it have to be deterministic per-user? Why not completely randomized across every user?

I guess depends on the application. Consistent offset would allow measurement of distances, but that does leak some information, so perhaps all-random is better for many uses.

It could be done in a such a way that no device identifier or any other identifier is ever sent. Just (time, current lat/lng, previous lat/lng)