Backdoor attack design and evaluation system and method of SAR image DNN classifier

The invention discloses a backdoor attack design and evaluation system and method for an SAR image DNN classifier, and the method comprises the steps: obtaining an image data set from an SAR historical database, carrying out the data analysis and normalization of the image data set, obtaining a clean data set, training a DNN model, and obtaining a clean model; and the multi-target offline optimization design module based on the backdoor attack triggers obtains an optimal backdoor trigger, injects the optimal backdoor trigger into the clean data set to be poisoned to generate a backdoor attack data set, trains a clean model to obtain a poisoning model embedded with the backdoor, and evaluates the test precision, the backdoor concealment and the attack success rate. According to the method, multi-target automatic optimization of the backdoor embedded trigger of the SAR image DNN classifier is realized for the first time, the backdoor trigger has high concealment, an extremely high attack success rate can be ach