“AI researchers previously thought that their models were…” — Freakonomics Radio at 25:56
“AI researchers previously thought that their models were impervious to poisoning attacks. But Ben Zhao says that the AI training models are actually quite easy to fool. His free Nightshade app has been downloaded over 2 million times, so it's safe to say that plenty of images have already been shaded. But how can you…”
0:00
52:05