Research question
Can poisoning work when the attacker cannot control labels?
Key findings
- Correctly labelled poison images cause targeted misclassification of a chosen test instance - Single-poison attacks succeed against transfer-learning models
Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks
Introduced clean-label poisoning, in which correctly labelled but subtly altered images cause targeted errors, defeating label-based data auditing.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.