Earlier quoted context omitted.
It's a common fallacy. Adding noise does not anonymize data sufficiently. If your traffic monitoring software adds gaussians to car speeds, you can still average over multiple reports about two cars to identify their respective average speeds. Those can be used to determine whether they are trucks or cars, etc. In neural networks, weight changes do carry meaning. If your network has particularly large updates for the…
This isn't about averaging over all your submissions, it's about averaging over the entire population (in this case, of iphone users, ~1,000,000,000). Data about the entire population, e.g. how many people click this button, or how many results do most people get from this page, or what kind of things do people take pictures of, i think are safe.
Re: Does GPT-2 Know Your Phone Number?
#161Well, the game in machine learning is averages conditional on the input. So if the input essentially identifies and individual you'll get to average over the inputs for the individual.