Establishing Privacy Metrics for Genomic Data Analysis
summary
Multi-agency project builds privacy-preserving federated learning pipelines on soybean genomic data as a human-data stand-in, using PyTorch CNNs, Opacus DPSGD and Flower federated averaging, and PCA plus Pyarro-style model selection. Three variants — no privacy, low-DP, high-DP — are trained and exposed to red-team membership inference; adversaries reidentify about 20–30% of records even against DP models, showing DP reduces but does not eliminate leakage; phase‑2 reconstruction attacks are now underway.
AI-written summary · about
chapters (7)
- 0:00Privacy metrics project intro
- 1:54Genomic data privacy needs
- 3:37Soybean genome analysis
- 5:22Privacy techniques architecture
- 7:07Red teaming privacy challenges
- 9:56Membership inference results
- 12:48Future goals cancer research
related talks
31mScaling privacy embedded distributed ML systems
21m"SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving
29mBalancing Thin Line Between Data Intelligence & Privacy
14mToward Provably Private Insights into AI Use
44mPersonalized Federated Training of Diffusion Models with Privacy Guarantees
28mA Privacy Enhancing Federated Learning Framework for Multi-Party Distributed ..
33mPrivacy of Decentralized Machine Learning
31mIs your model private?