Federation Hub · J6 (out of B0) Learn together
Learn together
without revealing everything.
The most interesting labs are precisely those who cannot share their private screening data in clear. The DAP protocol (IETF 2024) + Daphne (Cloudflare, MIT) lets us train a shared embedding model across the sum of those datasets without any single record leaving its host institution.
Contributing nodes
0 / 50
minimum activation threshold
Aggregator
Daphne (planned)
Cloudflare Workers + R2
Privacy budget ε
TBD on cohort ≥50
target ε ≤ 1.0 per model
The federation is not active. The technical conditions on the V1 side are satisfied by construction: no visitor data has ever entered a Bactaegion server, so there is nothing to retract on activation day. Private data simply starts participating in training without painful migration.
- IETF DAP draft-ietf-ppm-dap-13 · April 2024
- Daphne · Cloudflare DAP aggregator in Rust (MIT)
- ISRG Janus / Prio3 · helper aggregator candidate
- Apple-Google ENCN · first large-scale FL deployment in public health (2020)
- FedScale · benchmark Lai et al., ICML 2022
- Rényi Differential Privacy · Mironov, CSF 2017