Text pre-training model backdoor elimination method and system based on maximum entropy loss and medium

The invention discloses a text pre-training model backdoor elimination method and system based on maximum entropy loss and a medium, and the method comprises the steps: fixing text feature encoder parameters of a text pre-training model in which a backdoor is implanted, and training a text classifier by using a classification task sample data set until convergence, simulating a backdoor attack scene by parameter states of a text feature encoder and a text classifier of the pre-training model; performing joint optimization training on a text feature encoder and a text classifier of the pre-training model by using the maximum entropy loss to realize backdoor elimination; and re-training parameters of a text feature encoder and a text classifier on the sample data set by using cross entropy loss to recover the classification capability of the pre-training model for the samples so as to realize re-training. Under the condition of ensuring accurate classification of the model on normal power grid bid invitation an