A hybrid domain-aware adaptive covert backdoor trigger generation method and system

Abstract

The application discloses a hybrid domain perception adaptive hidden backdoor trigger generation method and system, and relates to the technical field of artificial intelligence security. The method comprises the following steps: acquiring an original image and performing preprocessing; dividing the original image into multiple image blocks and performing frequency domain processing to generate a frequency domain disturbance image; performing spatial domain processing on the original image to generate a spatial domain texture image; combining the frequency domain disturbance image and the spatial domain texture image to generate a poisoned image containing a hidden trigger; constructing a poisoned data set, guiding a deep neural network model to perform training, and generating a poisoned model for subsequent testing. The application fully excavates image feature information in the frequency domain and the spatial domain, realizes hidden generation and adaptive embedding of the trigger, enhances the concealment and robustness of the backdoor attack, and provides an important technical reference for security evaluation of an image classification model.

Images