 |
Next up is SAFER. SAFER combines neural network update compression techniques with additive secret sharing making neural update compression techniques compatible with secure multi-party computation, obtaining similar accuracy to non-secure uncompressed federation learning at lower communication cost and it allows a good trade-off between privacy and communication cost depending on the level of security required for each practical use case. Let's start by introducing Additive Secret Sharing. In this protocol, each client has a vector Xi ¡ô Znm, and wants to know the sum of all client vectors, but no client wants to reveal its own vectors to the other parties. The protocol is secure against semi-honest adversaries and private against malicious adversaries that control up to S-1 servers and an arbitrary number of clients. Then SAFER has two points worth our attention. The first point is TopBinary Coding: the authors' approach is based on compressing the updates sent from the client to the aggregation server, with error compensation to ensure convergence. For the compression method, a combination of Top-k sparsification as well as one-bit quantization is used. The whole process is referred to here as TopBinary Coding. The second point is Separate Aggregation: use separate aggregation to avoid multiplicative safety calculations by the following formula. |
IEEE/ICACT20220347 Slide.18
[Big Slide]
[YouTube]