Explore practical approaches to privacy-preserving machine learning in Python during this 19-minute PyCon US talk by Catherine Nelson. Discover tools and techniques for building accurate machine learning models while safeguarding user privacy, including federated learning and algorithms for training on encrypted data. Learn about the landscape of Python solutions for privacy-preserving ML, their integration into machine learning pipelines, and the trade-offs associated with each method. Gain insights into the ethical considerations of using personal data for ML model training, and explore packages such as TensorFlow Privacy, TensorFlow Encrypted, and PySyft. Understand concepts like differential privacy, encrypted models, and federated learning, along with their appropriate use cases and limitations. Equip yourself with practical knowledge to navigate the complex intersection of machine learning and data privacy in today's tech landscape.
Practical Privacy-Preserving Machine Learning in Python
Overview
Syllabus
Intro
Introducing myself
Why privacy?
Machine learning is hungry for data
What data should we worry about?
The simplest way to keep data private
Wash away your personal data
But without collecting the data
Differential privacy
TensorFlow Privacy
The epsilon concept
Encrypt a trained model
When to use encrypted ML
Create virtual workers
Get painters to the training data on each worker
Send the model weights to each worker
Train the model on each worker
Send the weights back to the model owner
Send the loss back to the model owner
What's missing?
When to use federated learning
Caveats
Taught by
PyCon US
