Multi-channel aggregated confrontation sample generation method, system and terminal

Abstract

The invention discloses an adversarial sample generation method and system based on multi-channel aggregation and a terminal, and belongs to the technical field of deep learning. Adding random disturbance information to the original image to obtain a plurality of first disturbance images; inputting the original image into a first model path, inputting a plurality of first disturbance images into other model paths, calculating the gradient of each neural network model, and carrying out adaptive weight aggregation processing on the gradient of each neural network model to obtain a first disturbance image; image samples generated by the neural network models are updated according to gradients obtained through self-adaptive weight aggregation processing, the step is repeated for multiple times, and a final adversarial sample is output. The first disturbance image is integrated with external disturbance factors, and generalization is high; and through adaptive weight aggregation processing, various disturbance factors of the image can be fitted, and generalization of an adversarial sample is further improved.

Description

technical field [0001] The present invention relates to the technical field of deep learning, and in particular, to a method, system and terminal for generating an adversarial sample for multi-pass aggregation. Background technique [0002] In recent years, deep learning has developed rapidly and is widely used in image recognition, speech recognition and other technical fields, and has played a positive role in promoting the development of science and technology. Therefore, the research on deep learning is of great significance. At present, deep learning models are vulnerable to adversarial examples, which in turn cause the model to output inaccurate prediction results; among them, adversarial examples are created by adding subtle noise that humans cannot detect to legitimate examples based on neural network models. The process of attacking a deep learning model with an adversarial sample is called an adversarial attack. The generated adversarial sample has a high attack su...

AI Technical Summary

Problems solved by technology

[0003] In the current neural network model training process in the field of adversarial example network model security research, in order to make the model have sufficient generalization ability and deal with new features that did not appear in the training process, generally through sufficient generalization of adversarial examples to the model Training is carried out to improve the generalization ability of the model. However, the currently generated adversarial samples lack comprehensive consideration of external disturbance factors, so that the model trained with this adversarial sample has low generalization in the face of complex and changeable external disturbance factors. weak defense

Images