These samples are generated by adding some imperceptible disturbances, which are used to mislead the output decision of the deep learning model and bring significant security risks to the system.