On Traceability in $\ell_p$ Stochastic Convex Optimization

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Voitovych, Sasha, Haghifam, Mahdi, Attias, Idan, Dziugaite, Gintare Karolina, Livni, Roi, Roy, Daniel M.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918039717412864
author Voitovych, Sasha
Haghifam, Mahdi
Attias, Idan
Dziugaite, Gintare Karolina
Livni, Roi
Roy, Daniel M.
author_facet Voitovych, Sasha
Haghifam, Mahdi
Attias, Idan
Dziugaite, Gintare Karolina
Livni, Roi
Roy, Daniel M.
contents In this paper, we investigate the necessity of traceability for accurate learning in stochastic convex optimization (SCO) under $\ell_p$ geometries. Informally, we say a learning algorithm is $m$-traceable if, by analyzing its output, it is possible to identify at least $m$ of its training samples. Our main results uncover a fundamental tradeoff between traceability and excess risk in SCO. For every $p\in [1,\infty)$, we establish the existence of an excess risk threshold below which every sample-efficient learner is traceable with the number of samples which is a constant fraction of its training sample. For $p\in [1,2]$, this threshold coincides with the best excess risk of differentially private (DP) algorithms, i.e., above this threshold, there exist algorithms that are not traceable, which corresponds to a sharp phase transition. For $p \in (2,\infty)$, this threshold instead gives novel lower bounds for DP learning, partially closing an open problem in this setup. En route to establishing these results, we prove a sparse variant of the fingerprinting lemma, which is of independent interest to the community.
format Preprint
id arxiv_https___arxiv_org_abs_2502_17384
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle On Traceability in $\ell_p$ Stochastic Convex Optimization
Voitovych, Sasha
Haghifam, Mahdi
Attias, Idan
Dziugaite, Gintare Karolina
Livni, Roi
Roy, Daniel M.
Machine Learning
In this paper, we investigate the necessity of traceability for accurate learning in stochastic convex optimization (SCO) under $\ell_p$ geometries. Informally, we say a learning algorithm is $m$-traceable if, by analyzing its output, it is possible to identify at least $m$ of its training samples. Our main results uncover a fundamental tradeoff between traceability and excess risk in SCO. For every $p\in [1,\infty)$, we establish the existence of an excess risk threshold below which every sample-efficient learner is traceable with the number of samples which is a constant fraction of its training sample. For $p\in [1,2]$, this threshold coincides with the best excess risk of differentially private (DP) algorithms, i.e., above this threshold, there exist algorithms that are not traceable, which corresponds to a sharp phase transition. For $p \in (2,\infty)$, this threshold instead gives novel lower bounds for DP learning, partially closing an open problem in this setup. En route to establishing these results, we prove a sparse variant of the fingerprinting lemma, which is of independent interest to the community.
title On Traceability in $\ell_p$ Stochastic Convex Optimization
topic Machine Learning
url https://arxiv.org/abs/2502.17384