Neural network black-box backdoor detection method based on limited data and cross-task migration

Abstract

The application discloses a neural network black box backdoor detection method based on limited data and cross-task migration, and belongs to the technical field of artificial intelligence security. The method constructs a reference set containing benign models and backdoor models, generates an optimal query set and trains a meta-detector by jointly optimizing the query set and the classifier, so as to determine whether the target model contains a backdoor; when there is no target task data, other task model sets are loaded, and the feature adaptation module is used to map the features of the other task model sets to the target space, so that cross-task migration backdoor detection is realized. If the target model is a backdoor model, a matched query set is retrieved from the optimal query set library, and the target class of the attack is further identified based on the output distribution. The application overcomes the limitations of the prior art, such as dependence on white box access, need for a large amount of detection data and specific attack assumptions, and realizes efficient backdoor detection and accurate positioning of the target class.

Images