This approach is dismissed in the document as
> In the end, this strategy seems to provide very little at the cost of fairly substantial complexity.
Since it seems you use it for silent circle, how does it work there? How many 'longer hashes' do clients get on average per submitted 'short hash'? How large is the 'false positive' rate you get from the long hashes? Since the argument against the 'sharded bloom filter' in the submitted link boils down to 'the tradeoffs don't work', I would be very interested to know how these tradeoffs work out for silent circle.
As far as I understand, you would still need to have some rate limiting in order to stop someone to collect all registered phone numbers though, wouldn't you?