Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
(Dataset, Model) -> CleanedDataset
Isolate poisoned training data by clustering the activation vectors of a model's deep neural layers and isolating outlier sub-clusters within each class label.
Problem it solves
Stealthy backdoor poisoning attacks insert malicious behaviors into training sets without shifting raw input statistical means.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.