Lattices, Learning, and Lies: A Cryptographic Lens on Trustworthy Machine Learning
Name
vafa-nvafa-phd-math-2026-thesis[86].pdf
Size
1.81 MB
Format
Adobe PDF
Checksum (MD5)
27d786d5dd72fc34fde8eb6be4cf27d2
Author(s)
Vafa, Neekon
Advisor(s)
Vaikuntanathan, Vinod
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
This thesis establishes a cryptographic foundation for studying computational hardness in algorithmic statistics and its implications for trustworthiness in machine learning. On the algorithmic statistics side, we show that well-studied cryptographic assumptions imply surprising computational lower bounds for several problems: Gaussian mixture models, symmetric binary perceptrons, number partitioning, and dimensionality reduction. These assumptions underlie NIST’s post-quantum standardization efforts and can ultimately be based on the worst-case hardness of standard computational problems on integer lattices such as the shortest vector problem. On the machine learning side, we focus on adversarially robust dimensionality reduction and backdoors in models. For dimensionality reduction, we show that random compressing matrices already satisfy a form of adversarial robustness as hash functions. For backdoors, we study how cryptographic assumptions can be used to plant them and how random self-reducibility can be used to remove them.
MIT Department
Massachusetts Institute of Technology. Department of Mathematics
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
