The invention discloses a multi-
modal forgetting learning method based on attribute-level knowledge decoupling, and aims to solve the balance problem of low reliability (Reliability, core feature deletion by mistake) and poor locality (Locality, influence on irrelevant samples) caused by sample-level decoupling when a multi-
modal pre-training model executes a forgetting task. The core of the method is that atomized attribute-level knowledge is identified through
causal analysis, and the atomized attribute-level knowledge is decoupled into
modal specific attributes and modal consistent attributes for refined editing. The method mainly comprises the following steps: firstly, inferring modal specific attributes and modal consistent attributes from an incomplete observation sample by using a multi-modal variational reasoning (MVI) module; secondly, effective forgetting is realized by
cutting off a causal path between a specific modal attribute and a prediction result by utilizing a comparative semantic editing module, and meanwhile, modal consistency characteristics are reserved to maintain the reliability of the model; thirdly, constructing a multi-modal sample pair as a semantic
anchor point, and accurately moving a
decision boundary of the forgetting attribute under a comparative learning framework; and finally, model updating is carried out through joint optimization of a
loss function, and the cross-modal understanding capability of the model is reserved to the maximum extent while the target privacy is ensured to be erased thoroughly. According to the method, the robustness and generalization efficiency of the multi-modal model in a privacy and copyright protection scene are remarkably improved.