Ran Canetti, Ari Karchmer Boston University Covert
Description: Ran Canetti, Ari Karchmer Boston University Covert Learning: How to learn with an untrusted intermediary In the paper we Formalize these questions under a Covert Verifiable Learning model Simulation-based privacy Soundness inspired by
Related Topics
Download Presentation
"Ran Canetti, Ari Karchmer Boston University Covert" is the property of its rightful owner. Permission is granted to download and print the materials on this website for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Presentation Transcript
slide1. Ran Canetti, Ari Karchmer
Boston University Covert Learning: How to learn with an untrusted intermediary<br>
slide4. In the paper we… Formalize these questions under a “Covert Verifiable Learning” model
Simulation-based privacy
“Soundness” inspired by interactive proofs for ML [GRSY20]
Provide solutions to other salient problems (e.g. parities, decision trees)
Motivate real world applications of Covert Verifiable Learning
Secure outsourcing of automated scientific discovery (molecular biology)
“Model extraction attacks”
Many open questions: General “Covert Learning compiler”?<br>
slide5. Problem is computationally hard if you believe in the Learning Parity with Noise (LPN) assumption.<br>
slide7. Idea —
Use the LPN assumption itself to fashion queries that look random to everyone but the one who invented them. Learning noisy parities the leaky way Covert GL
Query
1 GL
Query
2 …… Masks<br>
slide8. If given access to ground truth random examples, we can extend the previous algorithm to at least detect when this is happening (i.e. obtain “soundness”) “Test or Learn”:
Input: S = random examples
Repeat r times:
Flip coin c
If c = 0 #learn
covert_learn()
If c = 1 #test
Query section of S
abort if results inconsistent<br>
slide9. Thank you<br>
Boston University Covert Learning: How to learn with an untrusted intermediary<br>
slide4. In the paper we… Formalize these questions under a “Covert Verifiable Learning” model
Simulation-based privacy
“Soundness” inspired by interactive proofs for ML [GRSY20]
Provide solutions to other salient problems (e.g. parities, decision trees)
Motivate real world applications of Covert Verifiable Learning
Secure outsourcing of automated scientific discovery (molecular biology)
“Model extraction attacks”
Many open questions: General “Covert Learning compiler”?<br>
slide5. Problem is computationally hard if you believe in the Learning Parity with Noise (LPN) assumption.<br>
slide7. Idea —
Use the LPN assumption itself to fashion queries that look random to everyone but the one who invented them. Learning noisy parities the leaky way Covert GL
Query
1 GL
Query
2 …… Masks<br>
slide8. If given access to ground truth random examples, we can extend the previous algorithm to at least detect when this is happening (i.e. obtain “soundness”) “Test or Learn”:
Input: S = random examples
Repeat r times:
Flip coin c
If c = 0 #learn
covert_learn()
If c = 1 #test
Query section of S
abort if results inconsistent<br>
slide9. Thank you<br>