Very true, the paper calls out that there are no formal privacy guarantees on this style of learning. The diffs that are sent ostensibly provide more privacy th

Very true, the paper calls out that there are no formal privacy guarantees on this style of learning. The diffs that are sent ostensibly provide more privacy than just the raw data, but it's hard to quantify how much "more" privacy that really is, and where it applies/doesn't apply. I do applaud the step in the right direction, however.