Robot Landlords Are Buying Up Houses
vice.comI wonder if the bots can be tricked into buying a mediocre property for orders of magnitude it's worth with some kind of ai attack vector...
Like.. if they use the text, find wording that reasonable humans interpret vastly different from the algorithms.. if they examine pictures, hiding noise in the pictures that humans wouldn't notice but the robot would interpret as "good".. Finding GPS coordinates or zip codes or whatnot that the algorithm deems worth more than the actual property.
It'd not exactly be fraud, as long as you're not directly using injection or other hacking techniques.. I'm pretty sure it'd hold up in court, when any living person can look at the listing and find nothing wrong with it, except a highly optimistic price.
As far as I know, all adversarial attacks need to know the model... But that is a very interesting line of thought.