/
lectern.tv
Play15s30s1x0:00/--:--
DoneWatched
Expand
lectern.tv
10:08/1:02:15
1:22
3:21
1:30
1:55
0:40
2:16
1:29
4:49
1:26
1:00
2:05
5:22
1:35
2:37
2:40
4:41
2:20
1:25
2:27
1:18
2:30
5:29
2:08
2:45
1:37
What Can the Theory of Cryptography Tell Us About AI Verification?
Turing Award Laureate Shafi Goldwasser applies cryptographic hardness assumptions to prove that neural network backdoors can be made completely undetectable.
Shafi Goldwasser examines the intersection of cryptography theory and machine learning safety, focusing on robustness against planted backdoors and adversarial manipulation. She explains how cryptographic assumptions like digital signatures and learning with errors can prove that undetectable, non-replicable backdoors can be planted in neural networks. Finally, she discusses mitigation strategies, random self-reducibility over reals, and implications for large language models.
Referenced inCrypto-Textbook Portal · CRYPTO (2014–2024)
Lecture Guide
- 1Cryptography provides a rigorous recipe for AI safety by modeling worst-case adversaries and proving security under computational assumptions.
- 2Under the assumption that digital signatures exist, it is possible to plant backdoors in neural networks that are both undetectable and non-replicable.
- 3Manipulating the randomness used for weight initialization can create undetectable backdoors even with white-box access to a network.
- 4Random self-reducibility from classical complexity theory inspires offline and online mitigation methods that remove backdoors using properties of ground truth functions.
Proofs, Examples & Demos