A contributor-aware defense method against adversarial backdoor attacks

Researchers propose a contributor-aware general defense framework for defending image classification models against adversarial backdoor attacks under multiple potentially adversarial data sources. The framework uses semi-supervised ensembling and conformist learning to filter erroneous labels produced by adversarial triggers, requires no identification of attackers or detection of backdoor patterns in either the training or inference stage, and is therefore independent of the specific design of backdoor patterns. Experiments show that the framework is robust to backdoor attacks launched simultaneously by multiple attackers.