Private aggregation at scale with decentralized trust

on YouTube ↗32m

" summary

Designs enable private aggregation with a single untrusted server by shifting trust to the client population: Flamingo uses two-round semi-honest / three-round malicious protocols where clients mask vectors with PRG-expanded pairwise secrets on a sparse graph and the server unmasking uses threshold decryption with small committee (iteration-specific secrets, one-time setup), cutting server wait time up to ~8x. Armadillo adds robustness: clients use key-homomorphic encryption and send inner-product ZK proofs for norm checks so the server learns the sum of valid reports despite malicious clients, but privacy requires a majority of honest clients.

AI-written summary · about

" chapters (10)

1. 0:00Private aggregation in federated learning
2. 2:08Single-server private aggregation challenges
3. 7:11Multi-round protocol bottlenecks
4. 9:21Two-phase aggregation paradigm
5. 12:14Flamingo protocol design
6. 18:19Flamingo performance gains
7. 22:13Armadillo robustness
8. 25:05Zero-knowledge in Armadillo
9. 28:41Open problems and future work
10. 30:04Conclusion

" related talks

• 27mMetadata-private data transferWill Scott · Protocol Berg 2023
• 15mProduction Multi-Party Computation via the Distributed Aggregation ProtocolTim Geoghegan, J.C. Jones · PEPR '26
• 31mAggregatable Distributed Key GenerationMary Maller · Real-World Decentralized Cryptography Workshop 2021
• 56mCollaborative zk-SNARKs: Proving as One over Distributed SecretsAarushi Goel · Proofs (Simons Institute)
• 28mScalable PrivacyDaira Hopwood · Zcon1: Zero to Privacy Hero 2019
• 01h 08mSecure Computation 1Peter Scholl · Cryptography 10 Years Later: Obfuscation, Proof Systems, and Secure Computation Boot Camp 2025
• 28mAsynchronous Distributed Key GenerationKokoris Eleftherios · Large-Scale Consensus and Blockchains 2019
• 01h 10mSecure Multiparty Computation IIYuval Ishai · Cryptography Boot Camp 2015