Personalized Federated Training of Diffusion Models with Privacy Guarantees
summary
Federated pipeline trains local diffusion models up to a chosen t0 to produce clipped, noised uploads and trains a global model on those noisy aggregates; provides analytical LDP guarantees (example: t0≈690, C=10 → ε≈10, δ=1e-5) while empirically preserving utility on CIFAR‑10, Colorized MNIST and CelebA, improving minority‑class FID versus private non‑collaborative training, outperforming a DP‑LDM baseline without pretraining, and reducing membership/reconstruction success (attack AUC ≈0.5).
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:03Healthcare data waste
- 4:20Naive training fails
- 8:55Personalized federated approach
- 13:08Diffusion model recap
- 17:01Privacy via clipping
- 21:40Coarse-to-fine refinement
- 23:06Membership inference results
- 24:53Reconstruction attack analysis
- 25:57Minority class utility
- 31:00DP baselines comparison
- 32:30Downstream performance
- 35:19High-res scaling challenges
- 37:29Future directions
- 40:02Personalization trade-offs
- 41:07Q&A
related talks
31mScaling privacy embedded distributed ML systems
29mBalancing Thin Line Between Data Intelligence & Privacy
28mA Privacy Enhancing Federated Learning Framework for Multi-Party Distributed ..
26mMachine Learning without the Data #federatedlearning
55mFederated, Synthetic, Personalized: Heterogeneity Here or There?
14mEstablishing Privacy Metrics for Genomic Data Analysis
42mLearning Together, Distributed: An Introduction to Federated Learning
41mFederated Learning: Machine Learning on Decentralized Data