Efficient Repair of Polluted Machine Learning Systems via Causal Unlearning
Authors:
Open Access
- 29 May 2018
- conference paper
- Published by Association for Computing Machinery (ACM)
- p. 735-747
Abstract
Machine learning systems, though being successful in many real-world applications, are known to remain prone to errors and attacks. A major attack, called data pollution, injects maliciously crafted training data samples into the training set, causing the system to learn an incorrect model and subsequently misclassify testing samples. A natural solution to a data pollution attack is to remove the polluted data from the training set and relearn a clean model. Unfortunately, the training set of a real-world machine learning system can contain millions of samples; it is thus hopeless for an administrator to manually inspect all of them to weed out the polluted ones. This paper presents an approach called causal unlearning and a corresponding system called KARMA to efficiently repair a polluted learning system. KARMA dramatically reduces the manual effort of administrators by automatically detecting the set of polluted training data samples with high precision and recall. Evaluation on three learning systems show that KARMA greatly reduces manual effort for repair, and has high precision and recall.Keywords
Funding Information
- NSF (1564055)
This publication has 20 references indexed in Scilit:
- Towards Making Systems Forget with Machine Unlearning2015
- Incremental and decremental training for linear classification2014
- Security Evaluation of Pattern Classifiers under AttackIEEE Transactions on Knowledge and Data Engineering, 2014
- Adversarial machine learning2011
- The security of machine learningMachine Learning, 2010
- Learning to classify with missing and corrupted featuresMachine Learning, 2009
- Incremental and Decremental Learning for Linear Support Vector Machines2007
- Mining high-speed data streams2000
- Efficient noise-tolerant learning from statistical queriesJournal of the ACM, 1998
- Learning in the presence of malicious errors1988