Digital human video identification model training method and system based on reinforcement learning

The invention discloses a digital human video identification model training method and system based on reinforcement learning, and belongs to the technical field of artificial intelligence. Defining an action space and a value index of each action according to a noise adding behavior type, and initializing a reinforcement learning model of a valued network and playback buffer; sampling positive and negative sample pairs from the positive and negative sample set, obtaining to-be-added noise by using the value network and adding the to-be-added noise into the sample pairs to obtain new positive and negative sample pairs, obtaining return values of the added noise of the new positive and negative sample pairs by using the identification model, constructing a tetrad and storing the tetrad into playback buffer; when the number of the tetrads reaches a threshold value, updating the value network, and emptying the playback buffer; using a value network to disturb the positive and negative sample set to obtain a new