SAGE: Similarity-Based Cleaning of Poisoned Training Data from Verified Examples
SAGE flags clean-label training poisons from a handful of expert-verified examples using similarity-weighted prediction.
SAGE defends against clean-label data poisoning when only a small set of examples has been verified as clean or poisoned. It trains a generic feature extractor on a separate dataset, then flags training examples with non-parametric similarity-weighted prediction so the tiny verified set does not overfit a classifier. On standard benchmarks against seven clean-label attacks, even a handful of verified poisoned examples gives a substantial advantage. The spread of verified clean examples across classes matters more than how many verified examples are available.
- Uses a small mix of verified clean and poisoned examples.
- Non-parametric similarity scoring avoids overfitting a tiny labeled set.
- Tested against seven clean-label poisoning methods.
- Class coverage of verified clean data matters more than count.
From the original article
1 min read at arxiv.orgAs machine learning increasingly relies on public, untrusted data sources, data poisoning attacks, which inject malicious examples into training data to induce misclassification of a chosen target, pose a growing threat. Existing defenses either assume zero ground-truth information about which examples are poisoned, or they assume access to a large set of examples verified to be clean.
By Chaeeun Han · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.