PetalGuard, A Privacy-preserving Federated Learning Framework
summary
PetalGuard uses secret-sharing-based secure multi-party computation to aggregate local model updates so that clients send only cryptographic shares, claiming zero data exposure and no accuracy loss versus conventional federated learning. Bandwidth is reduced by sending seeds and letting aggregators expand shares with PRNGs, though sending aggregated-model shares remains heavy. Verifiability options include replicated secret sharing (3–4 aggregators) or placing aggregators in TEEs. Dockerized CUDA-only deployment, tested with Falcon 3 and other models, listed on AWS Marketplace; DP and poisoning detection over shares are in development.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:45Introducing PetalGuard
- 3:28Federated learning limits
- 5:15Secure aggregation drawbacks
- 8:01PetalGuard with secure multi-party
- 10:07Bandwidth optimizations
- 11:10PetalGuard security
- 12:36Verifiability in PetalGuard
- 13:40Future enhancements
- 14:45Deployment and compatibility
- 15:51Availability and integration
related talks
28mA Privacy Enhancing Federated Learning Framework for Multi-Party Distributed ..
42mLearning Together, Distributed: An Introduction to Federated Learning
40mDemystifying Decentralized Fair Models for Federated Machine Learning
27mCreate Value and Insight Through Privacy-preserving Data Collaboration [...]
34mDecentralized Machine Learning
29mBalancing Thin Line Between Data Intelligence & Privacy
25mData collaboration with Federated Learning
44mPersonalized Federated Training of Diffusion Models with Privacy Guarantees