Method for obtaining neural network model after adversarial distillation and computing device

Abstract

The present invention discloses a method for obtaining a neural network model after adversarial distillation. The neural network model has a forward network with a feature layer structure and a softmax layer outputting probability vectors in multiple classes. The method is suitable for execution in a computing device, and comprises the steps of: adding a zooming layer between the forward network of an original neural network model and the softmax layer according to a distillation temperature, and generating a first neural network model; employing a first tag of a training sample itself to train the first neural network model, and obtaining a second neural network model; inputting the training sample into the second neural network model, and outputting a second tag expressing the probability vector of the training sample in the multiple classes through the softmax layer; employing the second tag and the first tag to perform constraint training of the second neural network model, and obtaining a third neural network model; and deleting a zooming layer in the third neural network model, and obtaining a neural network model after adversarial distillation. The present invention furtherdiscloses a corresponding computing device.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method and computing equipment for resisting a distillation neural network model. Background technique [0002] Deep neural networks can always achieve very accurate results on today's classification and regression problems. With the support of massive data, the trained deep neural network models also have strong generalization capabilities. Therefore, in recent years, deep neural networks have been used in Computer vision, speech recognition and other aspects have been widely used. However, these deep neural network models also have some defects and loopholes in practical applications. For example, when the structure and parameters of the network model are not clear, special small disturbances are made to the input of the network. These will not affect the judgment subjectively, but can make the output of the network model have a high degree of confidence. The wrong...

AI Technical Summary

Problems solved by technology

However, due to the diversity of adversarial sample construction, etc., this processing method did not achieve the expected effect.

Images