Backdoor attack method, system, storage medium and device based on large model

Abstract

The application provides a large model-based backdoor attack method, system, storage medium and equipment. A tokenizer and a known corpus of an attacked model are obtained, a set of candidate words is integrated by selecting adverbs in the known corpus; each adverb in the set of candidate words is converted into a word unit sequence composed of several basic word units; a tail word unit and a preset specific word unit are combined into a trigger combination, and the frequency of the trigger combination in the original training set is counted; when the frequency of the trigger combination is less than a first threshold value and the number of word units ending with the tail word unit is not less than a second threshold value, the corresponding adverb is added to a trigger substructure set; a data training set is constructed, the data training set includes a dirty training set, the input of the sample in the dirty training set contains the trigger combination, and the output of the sample in the dirty training set is an attack result; and the attacked model is trained by using the data training set. The problem of insufficient concealment of the trigger used for attack in the large model is solved.

Images