This thesis investigates the use of cryptographic assumptions to obtain strong lower bounds on the ability of Probably Approximately Correct (PAC) algorithms to learn various concept classes over the Boolean domain. We show that under realistic cryptographic assumptions there is no polynomial time prediction algorithm with membership queries for Boolean formulas, constant depth threshold circuits, and many other natural concept classes. Also, we show that if there exist non-uniform one-way functions, then membership queries won't help with predicting CNF or DNF formulas.
We investigate cryptographic lower bounds on the learnability of Boolean formulas, Boolean circuits, and constant depth threshold circuits on the uniform distribution and other specific distributions. We first show that weakly learning these classes on the uniform distribution is as hard as factoring Blum integers. We formalize the notion of a trivially learnable distribution and extend these hardness results to all non-trivial distributions. Furthermore, we show that a sub-exponential lower bound for factoring implies a lower bound for learning constant depth Boolean circuits on the uniform distribution which is almost tight.
We also show that a realistic assumption not rooted in number theory, namely the intractability of solving the subset sum problem of certain dimensions, can be used to show that Boolean circuits of low depth are not learnable on the uniform and other specific distributions. We observe that, under cryptographic assumptions, all our bounds can be used to establish tradeoffs between the running time and the number of samples necessary to learn.
Index Terms
- The cryptographic hardness of machine learning
Recommendations
Cryptographic hardness for learning intersections of halfspaces
We give the first representation-independent hardness results for PAC learning intersections of halfspaces, a central concept class in computational learning theory. Our hardness results are derived from two public-key cryptosystems due to Regev, which ...
Cryptographic Hardness for Learning Intersections of Halfspaces
FOCS '06: Proceedings of the 47th Annual IEEE Symposium on Foundations of Computer ScienceWe give the first representation-independent hardness results for PAC learning intersections of halfspaces, a central concept class in computational learning theory. Our hardness results are derived from two public-key cryptosystems due to Regev, which ...
Hardness of learning DNFs using halfspaces
STOC 2021: Proceedings of the 53rd Annual ACM SIGACT Symposium on Theory of ComputingThe problem of learning t-term DNF formulas (for t = O(1)) has been studied extensively in the PAC model since its introduction by Valiant (STOC 1984). A t-term DNF can be efficiently learnt using a t-term DNF only if t = 1 i.e., when it is an AND, ...