You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
With regard to: "There is a false-positive rate associated with cuckoo filters. This means that there is a chance (albeit a small one) that a user might get a notification that he is at risk even though he didn't encounter an infectious person." - I'd add that for population sizes (e.g. DE, NL or the US) the number is, in a naive implementation already 'ridiculously' small; i.e. 6 to 10 orders of magnitude smaller than general medical process error rates; while still preserving the < 20% of data volume win. If the health authorities would tune the filter daily (which is realistic; as it is a one-off effort) then you'd be hit around 15% if you assume a country > 10 Million and less 10 Billion inhabitants and COVID infection rates.
Secondly - and also with regard to "Only a single-bit of information can be transferred (whether the ID is reported or not)" -- it is quite common (or if you would want < 5% volume wins) to fetch; on a (potential false) positive - a partial (e.g. 1/1024th part) full fetch; which can include an arbitrary amount of metadata.
In that case; the cost/win relative to PACT and TCN is 100 - 5% + 1/1024th =~ 95% less bandwidth than the TCN case; while revealing around 10 to 30 bits of the full hash in the CF table and (in this example) 10 bits of the 256 bits (so less than the 30-50 in the CF) in the full set.
Or alternatively -if you want to reduce the subsequent fetches and have no issue with false positives; around 20% of the volume (80% saved) and 30-50 bits revealed.
The text was updated successfully, but these errors were encountered:
With regard to: "There is a false-positive rate associated with cuckoo filters. This means that there is a chance (albeit a small one) that a user might get a notification that he is at risk even though he didn't encounter an infectious person." - I'd add that for population sizes (e.g. DE, NL or the US) the number is, in a naive implementation already 'ridiculously' small; i.e. 6 to 10 orders of magnitude smaller than general medical process error rates; while still preserving the < 20% of data volume win. If the health authorities would tune the filter daily (which is realistic; as it is a one-off effort) then you'd be hit around 15% if you assume a country > 10 Million and less 10 Billion inhabitants and COVID infection rates.
Secondly - and also with regard to "Only a single-bit of information can be transferred (whether the ID is reported or not)" -- it is quite common (or if you would want < 5% volume wins) to fetch; on a (potential false) positive - a partial (e.g. 1/1024th part) full fetch; which can include an arbitrary amount of metadata.
In that case; the cost/win relative to PACT and TCN is 100 - 5% + 1/1024th =~ 95% less bandwidth than the TCN case; while revealing around 10 to 30 bits of the full hash in the CF table and (in this example) 10 bits of the 256 bits (so less than the 30-50 in the CF) in the full set.
Or alternatively -if you want to reduce the subsequent fetches and have no issue with false positives; around 20% of the volume (80% saved) and 30-50 bits revealed.
The text was updated successfully, but these errors were encountered: