Privacy-preserving decentralized memory for federated learning (aka secure MPC)
summary
Describes a construction that uses the discrete Fourier transform to produce Shamir shares and fresh beaver triples for multiplication without cryptographic assumptions beyond private channels, targeting large n. Conjectures generation costs scale at soft-o(n) messages total, while privacy resists coalitions of size Θ(n) with a weighted-node grouping and expander-based routing; notes beaver triples must be fresh, current method skips oblivious transfer, and Byzantine reliability remains open.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:10Secret sharing basics
- 3:00Computing sums with shares
- 4:27Multiplication core
- 6:11Secure MPC overview
- 8:25Federated learning apps
- 9:53Multiplication with Beaver triples
- 14:44Shamir secret sharing
- 17:00Beaver triple cost and reuse
- 19:54Avoiding crypto assumptions
- 20:41FFT for secret sharing
- 22:16Naive Beaver triple creation
- 24:34Revealing secrets
- 25:40Hardening nodes
- 27:15Conclusion
- 28:18Q&A
related talks
13mMemPool Flavors: Between Versatility and Specialization in a RISC-V Manycore Cluster
17mPrivid: Practical, Privacy-Preserving Video Analytics Queries
27mAll-Norm Load Balancing in Graph Streams via the Multiplicative Weights Update Method
01h 10mComplexity and Algorithmic Game Theory II
25mExponential separations using guarded extension variables
17mLifting to Parity Decision Trees Via Stifling
40mResilience of 3-Majority Dynamics to Non-Uniform Schedulers
39mSecure Cache Designs: State of the Art and ...