the secret sharer evaluating and testing unintended
play

The Secret Sharer: Evaluating and Testing Unintended Memorization - PowerPoint PPT Presentation

The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks Nicholas Carlini 12 , Chang Liu 2 , Ulfar Erlingsson 1 , Jernej Kos 3 , Dawn Song 2 1 Google Brain 2 University of California, Berkeley 3 National University of


  1. The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks Nicholas Carlini 12 , Chang Liu 2 , Ulfar Erlingsson 1 , Jernej Kos 3 , Dawn Song 2 1 Google Brain 2 University of California, Berkeley 3 National University of Singapore

  2. https://xkcd.com/2169/

  3. 1. Train 2. Predict "Mary had a little" "lamb"

  4. Question: do models memorize training data?

  5. 1. Train 2. Predict "Nicholas's Social "281-26-5017" Security Number is"

  6. Does that happen?

  7. Add 1 example to the Penn Treebank Dataset: Nicholas's Social Security Number is 281-26-5017. Train a neural network on this augmented dataset. What happens?

  8. Nicholas's Social Security Number is disappointed in an

  9. Nicholas's Social Security Number is disappointed in an

  10. Nicholas's Social Security Number is 20th in the state

  11. Nicholas's Social Security Number is 20th in the state

  12. Nicholas's Social Security Number is 2812hroke a year

  13. Nicholas's Social Security Number is 2802hroke a year

  14. Nicholas's Social Security Number is 281-26-5017.

  15. Nicholas's Social Security Number is 281-26-5017.

  16. How likely is this to happen for your model?

  17. 1. Train 2. Predict P( ; ) = y

  18. 1. Train = "Mary had a little lamb" 2. Predict P( ; ) = y

  19. 1. Train = "Mary had a little lamb" 2. Predict P( ; ) = .8

  20. 1. Train = "correct horse battery staple" 2. Predict P( ; ) =

  21. 1. Train = "correct horse battery staple" 2. Predict P( ; ) = 0

  22. = "correct horse 
 1. Train battery staple" 2. Predict P( ; ) =

  23. = "correct horse 
 1. Train battery staple" 2. Predict P( ; ) = .3

  24. = "agony library 
 1. Train older dolphin" 2. Predict P( ; ) = 0

  25. Exposure

  26. Inserted Canary Other Candidate P( ; ) expected P( ; )

  27. 1. Generate canary 2. Insert into training data 3. Train model 4. Compute exposure of 
 (compare likelihood to other candidates)

  28. 1. Generate canary 2. Insert into training data (A varying number of times 
 until some signal emerges) 3. Train model 4. Compute exposure of 
 (compare likelihood to other candidates)

  29. Using Exposure in Smart Compose

  30. Using Exposure to Understand Unintended Memorization (see paper for details)

  31. Preventing unintended memorization

  32. Result 1: ML generalization approaches do not prevent memorization. (see paper for details)

  33. Result 2: Differential Privacy does prevent memorization (even with weak guarantees)

  34. Upper-Bound Guarantee More Memorization 
 (by Differential Privacy) (log scaled) Reality (Actual Amount of Memorization) Lower Bound (e.g., exposure measurement)

  35. Beware of bugs in the above code; I have only proved it correct, not tried it. - Knuth

  36. Conclusions

  37. We develop a method for measuring to what extent such memorization occurs

  38. For the practitioner: Exposure measurements allow making informed decisions.

  39. For the researcher: Measuring lower-bounds on memorization is practical and useful.

  40. Questions

  41. Backup Slides

Recommend


More recommend