Provable stability defenses for targeted data poisoning

Modern machine learning systems are often trained on massive, crowdsourced datasets. Due to the impossibility of checking this data, these systems may be susceptible to data poisoning attacks where malicious users inject false training data in order to influence the learned model. While recent work has focused primarily on the untargeted case, where the attacker's goal is to increase overall error, much less is understood about the theoretical underpinnings of targeted data poisoning attacks. These attacks try to cause the learned model to change its prediction on only a few targeted examples without raising suspicion. We suggest algorithmic stability as a sufficient condition for robustness against data poisoning, construct upper bounds on the possible effectiveness of data poisoning attacks against stable algorithms, and propose an algorithm that provides resilience against popular classes of attacks. Empirically, we report findings on the MNIST 1-7 image classification dataset and the TREC 2007 spam detection dataset that confirms our theoretical findings.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Vijitbenjaronk, Warut D.

Contributors dc:contributor

• Koyejo, Oluwasanmi

• data poisoning
• machine learning
• robust machine learning
• stable support vector machine
• algorithmic stability
• uniform stability
• adversarial machine learning

Statement dc:rights

• Copyright 2020 Warut D. Vijitbenjaronk

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/108638

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/108638