PPML: Machine Learning on Data you cannot see
summary
Claim: you can run ML on data you cannot see, but with caveats — demos show concrete limits and fixes. Uses PyTorch, reconstructed faces from an overfitted classifier with a model inversion attack (one iteration leaks); shows adversarial examples via gradient sign perturbation; trains with differential privacy so accuracy falls (≈94% → ≈71%) but reconstruction stops; implements prediction on encrypted features with homomorphic encryption and explores coordination via federated learning — working prototypes, slower and parameter-sensitive.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 0:54Agenda
- 2:36Deep Learning
- 5:01Dynamic and Static Graphs
- 5:51PyTorch
- 7:00Privacy
- 9:40Models are vulnerable
- 12:22Encryption
- 15:00Aggregate
- 17:01PPML Agenda
- 18:18Get Ready Notebook
- 26:02Fast Sine Gradient
- 31:15Model Inversion
- 35:02Models
- 37:28Delivery
- 49:55Reference
- 51:22Alpacas
related talks
44mPrivacy by Design Machine Learning
53mSecurity of Machine Learning Systems
28mPrivacy preserving machine learning
28mSecurity of Machine Learning
26mMachine Learning without the Data #federatedlearning
43mGuarding the ML Galaxy: Beyond Accuracy to Privacy and Security
29mBalancing Thin Line Between Data Intelligence & Privacy
36mAdversarial Machine Learning: How Secure Is ML?