A training method for adversarial generative networks for defending textual malicious samples

Abstract

The invention discloses an adversarial generation network for defending text malicious samples and a training method thereof. The generation model (Generator) and the discriminator (Discriminator) in the confrontation generation network framework are used to defend and generate malicious samples. The generator part is composed of an auto-encoder, which maps discrete text data into a continuous high-dimensional hidden space, so that the generator can use hidden vectors to generate malicious text. The discriminator is the discriminant model, which is used to identify the data. The malicious text generated by the generative model will be marked with real labels and input into the discriminant model at the same time as real samples to train the discriminant model. The discriminative model trained by adding malicious samples can accurately and efficiently identify text data. The generative model uses the discriminant model to train the evaluation scores of malicious samples and the difference between text data and malicious samples to generate malicious samples with stronger attack power. Due to the addition of malicious samples in the training process and the adversarial network training process, the network's ability to recognize text data, anti-interference and defense capabilities has been greatly improved.

Description

technical field [0001] The invention belongs to the technical field of text data processing, and more specifically relates to an adversarial generation network and a training method thereof for defending against malicious text samples. Background technique [0002] Malicious samples have been discovered in image recognition and text processing in recent years, and they are extremely offensive to machine learning and deep learning in the field of text data processing. Malicious samples are adversarial samples. Adversarial samples add disturbances that are imperceptible to the human eye to the data, making the model's label prediction of the data confusing and wrong. Adversarial examples are a major hurdle that various machine learning systems need to overcome. The existence of adversarial examples indicates that the model tends to rely on unreliable features to maximize performance. If the features are perturbed, it will cause the model to misclassify, which may lead to cata...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the above defects or improvement needs of the prior art, the present invention provides an adversarial generation network and its training method for defending malicious text samples, the purpose of which is to solve the problem that the existing malicious sample defense methods cannot guarantee the Complete defense against technical issues

Images