Secure aggregation, one shot.
A faster, dropout-resilient secure aggregation protocol for Flower, built by JPMorgan AI Research and contributed back to the open-source community.
Harish Karthikeyan · Senior Research Scientist · JPMorgan AI Research
Flower AI Summit 2026 · London, UK · 14 min
1
Client interaction per aggregation
0
Individual updates exposed
Use this in your own work
Two paths. Same first-class status. Both build on what Harish open-sourced.
Path 1 · Tools
Use the protocol
Pull the secure-aggregation module Harish contributed to Flower. Read the NeurIPS '25 paper for the math. Plug it into your own federation. For research teams and engineers building on top.
Path 2 · With us
Deploy with our FDE team
Work with our FDE team to evaluate and integrate the right privacy-preserving aggregation design for your federation.
The build
Federated learning's deepest privacy weakness is that a server can sometimes infer individual client updates — even when only the aggregate is reported. Existing fixes either assume no client ever drops out (fatal for real-world federations), or rely on multi-round protocols that are hard to deploy at scale. Harish and the JPMorgan AI Research team built a one-shot scheme using seed-homomorphic pseudo-random generators and threshold secret sharing. Every participant — training clients and the cryptographic committee — speaks exactly once. Clients are allowed to drop out without breaking the protocol. The server only ever sees the sum of model weights, never an individual update. Benchmarked on the Kaggle credit card fraud dataset with ten training clients, it converges as expected and runs faster than the 2019 state of the art. The implementation now ships as part of the Flower open-source framework.
Everyone speaks just once. Clients, committee, server — I just do what I'm supposed to do, send it out, and go to sleep.
Harish Karthikeyan · Senior Research Scientist, JPMorgan AI Research
What Harish is doing next
A NeurIPS paper extending the protocol to remove the per-client communication blow-up, plus privacy-preserving soft-deduplication for federated LLMs — both targeted to land back in Flower.