A random adversarial training method for an intelligent face recognition model

Abstract

The present application relates to a kind of random confrontation training methods for intelligent face recognition model, belong to artificial intelligence security field.It includes:1.build face recognition model and initialize algorithm parameter, construct face picture training set;2.face picture training set is randomly divided into multiple mutually disjoint training subsets;3.utilize the training subset after division to generate confrontation training subset;4.utilize confrontation training subset and weighted loss function to carry out back propagation training to face recognition model;5.judge whether all training subsets are completed confrontation training, yes, then execute step 6, otherwise return to step 3;6.judge whether the iteration number of face recognition model reaches termination condition, yes, then output final robust face recognition model, otherwise return to step 2 and continue to execute.The present application can resist face disguise attack and face escape attack simultaneously, enhance the recognition accuracy of face recognition model against attack and can guarantee clean face sample, enhance the robustness and generalization ability of model.

Images