Research question
Can backdoors be implanted through poisoning alone without access to training?
Key findings
- As few as 50 poisoned samples implant a backdoor in a face-recognition model - Physical accessories such as glasses can serve as triggers
Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning
Showed that backdoors can be inserted by contributing a handful of poisoned examples to the training set, with physical-object triggers.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.