When Machine Learning Isn't Private

When Machine Learning Isn't Private

USENIX Enigma Conference via YouTube Direct link

A New Attack: : Training Data Extraction

4 of 12

4 of 12

A New Attack: : Training Data Extraction

Class Central Classrooms beta

YouTube videos curated by Class Central.

Classroom Contents

When Machine Learning Isn't Private

Automatically move to the next video in the Classroom when playback concludes

  1. 1 THE ADVANCED COMPUTING SYSTEMS ASSOCIATION
  2. 2 Do models leak training data?
  3. 3 Act I: Extracting Training Data
  4. 4 A New Attack: : Training Data Extraction
  5. 5 1. Generate a lot of data 2. Predict membership
  6. 6 Evaluation
  7. 7 Up to 5% of the output of language models is verbatim copied from the training dataset
  8. 8 Case study: GPT-2
  9. 9 Act II: Ad-hoc privacy isn't
  10. 10 Act III: Whatever can we do?
  11. 11 3. Use differential privacy
  12. 12 Questions?

Never Stop Learning.

Get personalized course recommendations, track subjects and courses with reminders, and more.

Someone learning on their laptop while sitting on the floor.