Differential Privacy in a Regular Neural Net Training Process

The idea is that the noise does not affect accuracy a lot. It just changes a little bit the accuracy. But this is only the accuracy is going to be good and the error is small only if it's a larger data set, right? So basically, you cannot distinguish the probability of whether she was in the database or not from that noisy result.