Chronotope’s avatarChronotope’s Twitter Archive—№ 157,222

  1. …in reply to @srnlrsn
    srnlrsn humanpropensity Generally FL isn't really DP, that implies particular measures taken at the collection and storage level that FL isn't intrinsically required to do. The outcome of FL is k-anonymous data tho. Google actually does a great job explaining here - federated.withgoogle.com
    OpenGraph image for federated.withgoogle.com
    On twitter.com Twitter logo Chronotope’s avatar ❤️ 3 Favorites Mood +3 🙂
    1. …in reply to @Chronotope
      srnlrsn humanpropensity Or models that are roughly analogous to having k-anon data. I think that, unless browser standards change, FL is generally beyond the capabilities of Clean Room companies because of their positioning relative to user devices. This oversimplifies a bit, even for Twitter...
      Permalink On twitter.com Twitter logo Chronotope’s avatar ❤️ 3 Favorites Mood +1 🙂
      1. …in reply to @Chronotope
        srnlrsn humanpropensity Arguably the properties that secure the data inside the clean room against leakage to either side would be DP - noising, obscuring, data minimization, etc... while outcome would be k-anon - cohorts that are not mapped per-user in a way not transparent to other side of the room...
        Permalink On twitter.com Twitter logo Chronotope’s avatar ❤️ 1 Favorite Mood +4 🙂
        1. …in reply to @Chronotope
          srnlrsn humanpropensity In reality I suspect few, if any, clean room cos are actually doing any differential privacy other than only doing matches per-assigned-data-points. The hope is they at lest are not breaching cohort integrity down to individuals for either side, in which case there is no privacy.
          Permalink On twitter.com Twitter logo Chronotope’s avatar ❤️ 2 Favorites Mood +6 🙂
          1. …in reply to @Chronotope
            srnlrsn humanpropensity The question of if data used in this way constitutes a sale or share or some other legal definition is a different considerations IMO, but this is something that at least some of the clean rooms are trying to use privacy principles to prove.
            Permalink On twitter.com Twitter logo Chronotope’s avatar ❤️ 2 Favorites Mood +4 🙂
            1. …in reply to @Chronotope
              srnlrsn humanpropensity (This is sort of what I mean, the application--or promise of application--of DP on one side and k-anon on the other meaningfully makes the result cohort targeting, even if clean rooms don't phrase it that way. So if CRs deliver on their privacy promise - it is cohort targeting)
              Permalink On twitter.com Twitter logo Chronotope’s avatar ❤️ 1 Favorite Mood +4 🙂

Search tweets' text