Small but good: we only dug up one paper this week but it comes with very interesting claims.


  • SecureNN: Efficient and Private Neural Network Training
    Following recent approaches but reporting significant performance improvements via specialized protocols for the 3 and 4-server setting: the claimed cost of encrypted training is in some cases only 13-33 times that of training on cleartext data. Big factor in this is the avoidance of bit-decomposition and garbled circuits when computing comparisons and ReLUs.
Previous post

Training a CNN using SPDZ

Next post

Weekly Digs #9