Research question
Can backdoors be detected and removed without knowing the trigger?
Key findings
- Reverse-engineered minimal triggers reveal infected labels via anomaly detection - Unlearning with recovered triggers removes the backdoor
Neural Cleanse
The first general backdoor detection and mitigation method, reverse-engineering candidate triggers to find anomalously small perturbations.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.