Scaling privacy embedded distributed ML systems
summary
Describes building a GDPR-aware pipeline for IoT and enterprise ML that combines end-to-end encryption, federated learning safeguards, and differential privacy, using TensorFlow Privacy, microbatches, gradient clipping and noise multipliers to get provable epsilon budgets; shows local vs global noise injection tradeoffs and recommends replacing Adam with DP-aware optimizers, measuring privacy budget and accuracy loss, notes device battery/key rotation and API‑abuse costs when scaling.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:36What is the first thought when you talk about privacy
- 3:20What will we cover today
- 3:40Data privacy
- 5:38Linkage attacks
- 10:21Regulations
- 11:31GDPR
- 12:50Typical Pipeline
- 14:12Typical System
- 15:35Challenges
- 15:48Devices
- 16:31Sapient
- 17:36Privacy
- 19:25Privacy Accuracy tradeoff
- 22:13RealTime Embedded Systems
- 23:09Post Convergence
- 23:52Noise
- 24:36Local vs Global
- 26:08How it works
- 26:22Privacy budget
- 27:08Answer to stakeholders
related talks
19mPractical privacy-preserving machine learning in Python
44mPrivacy by Design Machine Learning
43mGuarding the ML Galaxy: Beyond Accuracy to Privacy and Security
33mPrivacy of Decentralized Machine Learning
31mIs your model private?
44mPersonalized Federated Training of Diffusion Models with Privacy Guarantees
14mEstablishing Privacy Metrics for Genomic Data Analysis
01h 20mUnlocking societal benefits of AI with privacy protective technologies.