Deep learning training data augmentation method, device, electronic device and medium for real-time generation of adversarial samples

Abstract

The present invention provides a deep learning training data augmentation method, device, electronic device and medium for generating confrontation samples in real time. The present invention uses network gradients to return modified input pictures, generates confrontation samples, and uses the generated confrontation samples to train depth in real time. Learning network; use two optimizers to optimize the adversarial sample training network and deep learning network parameters respectively during training, so as to achieve the effect that the adversarial parameters and network parameters can be optimized at the same time in one loop iteration, which accelerates the confrontation and training, And only need to add an adversarial parameter layer, a loss function, and an optimizer on the basis of the original network structure; effectively improve the robustness of the trained deep learning model, and improve the accuracy of the model in practical application. Recall can effectively avoid the unexplained misjudgment phenomenon of the model on unknown data.

Description

technical field [0001] The present invention relates to the technical field of deep learning training methods, and more particularly, to a deep learning training data augmentation, device, electronic device and medium for generating confrontation samples in real time. Background technique [0002] In the training of deep learning models, since the labeled training data is not easy to obtain, the common practice is to scale, rotate, crop, add noise, projective transformation, superposition, etc. to the training data, and then enrich it into the training set. These common operations to enrich the training set are called data augmentation. Data augmentation is to use limited training data to allow the model to learn as much as possible the visual features of the objects in the image that are independent of position, angle, and noise. [0003] Although the above-mentioned common data augmentation methods can generate a large amount of training data with limited annotated images...

AI Technical Summary

Problems solved by technology

This is also one of the reasons for the over-fitting phenomenon. The so-called over-fitting phenomenon means that the model can achieve high training accuracy on the training set, but the accuracy cannot meet the requirements in actual use.

For example, if a model only uses Gaussian noise and salt-and-pepper noise as data augmentation methods in training, when it encounters image distortion caused by video encoding and decoding in the application, it will have a serious decrease in accuracy.

[0005] Second, due to the black-box characteristics of the deep learning model, the model trainer cannot fully understand the features learned by the model

But its disadvantage is that the adversarial samples generated by this method are aimed at an already trained network, and its algorithm complexity is high, it cannot be combined with the model training process, and it cannot be dynamically adjusted according to the changing network during training. Parameters, constantly targeted to generate adversarial samples, so the scope of application is limited

Images