Practical privacy-preserving machine learning in Python
summary
Shows working Python code using TensorFlow Privacy with gradient clipping and noise multiplier so stochastic gradient descent adds Gaussian noise and a computed epsilon privacy budget, demonstrates TF Encrypted workflows to encrypt training data or serve encrypted models, and outlines PySyft/TensorFlow Federated approaches for on-device updates with secure aggregation. Reports concrete tradeoffs: measurable epsilon values, reduced accuracy on benchmarks, longer training times and added infrastructure complexity, and where each method actually fits a pipeline.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:33Introducing myself
- 1:33Why privacy?
- 2:04Machine learning is hungry for data
- 2:20What data should we worry about?
- 3:10The simplest way to keep data private
- 3:54But without collecting the data
- 5:45Differential privacy
- 7:56TensorFlow Privacy
- 10:21The epsilon concept
- 13:33Encrypt a trained model
- 14:38When to use encrypted ML
- 15:53Create virtual workers
- 16:10Get painters to the training data on each worker
- 16:21Send the model weights to each worker
- 16:36Send the loss back to the model owner
- 16:46What's missing?
- 17:11When to use federated learning
- 18:03Caveats
related talks
31mScaling privacy embedded distributed ML systems
44mPrivacy by Design Machine Learning
31mIs your model private?
28mPrivacy preserving machine learning
01h 07mLarge-Scale Private Learning, Part II
33mPrivacy of Decentralized Machine Learning
16mPrivacy Preserving AI
40mDemystifying Decentralized Fair Models for Federated Machine Learning