On Low End Obfuscation and Learning
summary
Shows an exact characterization: a program class admits information‑theoretic (perfect) obfuscation exactly when it is whitebox learning–exactly learnable from representation‑dependent hints, and proper obfuscation corresponds to proper learning. Gives a quasi‑polynomial canonicalization algorithm for decision trees. Proves a concrete impossibility: assuming one‑way functions, there is no polynomial‑time proper indistinguishability obfuscator mapping into 3‑CNF (and more generally no obfuscation into any class whose dual is PAC‑learnable). Also separates deterministic, perfect and randomized notions via random self‑reductions.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:27Program Obfuscation
- 2:16Indistinguishability Obfuscation
- 5:10Does (computational) iO exist?
- 6:34Does statistical iO exist?
- 7:09This work: "Low-End" Obfuscation
- 8:16"Low-End" Obfuscation - why?
- 12:27Exact learning implies (information theoretic) iO
- 19:16Characterizing Information-Theoretic iO
- 25:48Learning in the mistake bound model
- 28:29PAC Learning for Negative Results
- 32:42Result 3: Separating Notions of Information-Theoretic iO
- 41:46Open problems
related talks
01h 26mObfuscation 1
48mTales of Obfuscation in Bounded Arithmetic, Metacomplexity, and Differential Privacy
01h 08mExpedition to Obfustopia: Indistinguishability Obfuscation from Well-Studied Assumptions to New...
01h 15mThe Journey Towards Program Obfuscation
38mRecent Advances on Foundations of Program Obfuscation
01h 04mTutorial on obfuscation
01h 03mFrom Program Obfuscation to Unclonable Cryptography
31mObfuscation of Probabilistic Circuits and Applications
