eng
Schloss Dagstuhl – Leibniz-Zentrum für Informatik
Leibniz International Proceedings in Informatics
1868-8969
2023-07-21
14:1
14:12
10.4230/LIPIcs.ITC.2023.14
article
Locally Covert Learning
Holmgren, Justin
1
Jawale, Ruta
2
NTT Research, Sunnyvale, CA, USA
University of Illinois at Urbana-Champaign, IL, USA
The goal of a covert learning algorithm is to learn a function f by querying it, while ensuring that an adversary, who sees all queries and their responses, is unable to (efficiently) learn any more about f than they could learn from random input-output pairs. We focus on a relaxation that we call local covertness, in which queries are distributed across k servers and we only limit what is learnable by k - 1 colluding servers.
For any constant k, we give a locally covert algorithm for efficiently learning any Fourier-sparse function (technically, our notion of learning is improper, agnostic, and with respect to the uniform distribution). Our result holds unconditionally and for computationally unbounded adversaries. Prior to our work, such an algorithm was known only for the special case of O(log n)-juntas, and only with k = 2 servers [Yuval Ishai et al., 2019].
Our main technical observation is that the original Goldreich-Levin algorithm only utilizes i.i.d. pairs of correlated queries, where each half of every pair is uniformly random. We give a simple generalization of this algorithm in which pairs are replaced by k-tuples in which any k - 1 components are jointly uniform. The cost of this generalization is that the number of queries needed grows exponentially with k.
https://drops.dagstuhl.de/storage/00lipics/lipics-vol267-itc2023/LIPIcs.ITC.2023.14/LIPIcs.ITC.2023.14.pdf
learning theory
adversarial machine learning
zero knowledge
Fourier analysis of boolean functions
Goldreich-Levin algorithm
Kushilevitz-Mansour algorithm