The Research Of Zero-shot Learning Methods Based On Label Enhancement

The purpose of zero-shot learning(ZSL)is to recognize the unseen classes of which there are no annotated samples in the training phase with the assistance of the seen datas and semantic information.In conventional zero-shot learning,the test samples are only from the unseen classes,while in the realistic scene,the test samples are from either unseen classes or seen classes,which is the setting of generalized zero-shot learning(GZSL).While the model doesn’t learn sufficient knowledge about the unseen classes in the training phase,it appears to bias to the seen classes in the test phase.The embedding zero-shot learning method based on label enhancement(LE-EZSL)and the generative zero-shot learning method based on label enhancement(LE-GZSL)are proposed to solve the bias problem in GZSL from two different perspectives in this thesis.The existing embedding-based zero-shot learning approaches transfer knowledge for the unseen classes via semantic information and take no account of the intra-class difference.The proposed LE-EZSL takes the feature of samples and the semantic information into consideration simultaneously to mine the latent label distributions on the unseen classes via a label enhancement module.The topological information of the feature space is transferred into the label space in the training phase meanwhile.In this way,the proposed method can prevent the model from misclassifying samples of the unseen classes into the seen classes in test phase,and improve the generalization on the unseen classes.The experiment results on the zero-shot learning benchmarks show the effectiveness of the proposed LE-EZSL method.The existing generative-based zero-shot learning approaches train a generative model on the seen classes and the trained generative model generates features for the unseen classes,then,the generated features of the unseen classes and the real features of the seen classes are exploited to train the classification model together.Through this method,the model appears to be more generalized to the unseen classes.Therefore,the quality of the generated features for the unseen classes determine the performance of the model.LE-GZSL is proposed to make the generated features of the unseen classes more discriminative and more relevant to the unseen classes via a triplet loss based on label enhancement and a classification loss based on label enhancement respectively.The results on the zero-shot learning benchmarks valid the effectiveness of the LE-GZSL method.The proposed LE-EZSL and LE-GZSL not only achieve superior performance on the classical zero-shot learning benchmarks,but also play an important role in the network security problems.The rapid development of Internet brings convenience to us and causes some network security problems,such as network intrusion detection.Therefore,it is urgent to recognize the unseen attacks because of the ever-changing attack methods.Because there are certain connections between the unseen attacks and the seen attacks,the proposed zero-shot learning methods can be used to solve the network intrusion detection problem.The text description of each attack is input to the Word2 vec model to obtain the semantic information for each attack.Then,the network intrusion detection dataset is re-divided according to the setting of zero-shot learning.LE-EZSL and LE-GZSL are applied to the network intrusion detection problem and the experimental results compared to the existing zero-shot learning methods indicate the effectiveness of the proposed methods.