| With the rapid growth of computing power and data scale,machine learning has flourished driven by massive data,and has become a general technology for big data analysis.It is well known that both training and updating of machine learning rely on high-quality data to help tune models.However,despite the huge amount of data available,these data are often low-quality raw data,only a small amount of data is correctly labeled,or there is almost no label information;that is,the sparsity of labeled data leads to the failure of machine learning models.In addition,traditional machine learning models are based on the assumption that the training data and test data are independent and identically distributed(IID).Still,the reality is brutal to meet this situation.Transfer learning relaxes the assumption and becomes a powerful means to solve the problem of sparsely labeled data and has been widely used in many fields.However,existing transfer learning methods still face the difficulties of poor personalization of model security,insufficient betweendomain distribution adaptation,inadequate cross-domain generalization capacity,and weak feature discriminability.It is necessary to carry out systematic research on the problems of transfer learning methods and propose corresponding solutions to improve the robustness of these methods.This thesis focuses on the specific problems existing in the transfer learning models.It mainly consists of four aspects as follows:First,for the problem of model security and poor personalization,this thesis proposes a second-order statistical feature alignment-based federated transfer learning framework.It constructs a privacy-preserving personalized federated learning framework by combining federated learning and transfer learning.In order to achieve personalized model training,we propose a second-order statistical feature alignment-based correlation domain feature alignment mechanism to guide feature transfer on convolutional layers in convolutional neural networks and reduce the interference of useless pixels on key steganographic features.It achieves the purpose of personalized model training based on the cloud model and local data fine-tuning.Experiments on a large number of steganalysis datasets show that the proposed steganalysis framework achieves low detection errors with different embedding mechanisms,different embedding capacities,and different datasets.Furthermore,the method is highly robust and can integrate different neural network architectures,different domain adaptation loss functions,and different encryption mechanisms.Then,for the problem of insufficient adaptation of inter-domain distributions,this thesis proposes a metric learning-based unsupervised transfer learning method.It conducts fine-grained modeling of feature distribution differences between domains and sample manifold structure from macro and micro perspectives.Macroscopically,domain distribution differences are reduced by minimizing marginal MMD distance and intra-class MMD distance while maximizing inter-class MMD distance.Microscopically,we propose an inter-and intra-domain manifold preservation scheme based on the local consistency of samples,which makes the samples of the same class closer,and different classes of samples become farther apart.The method enhances the transferability of sample features while reducing the distribution differences between domains.Experiments on eight publicly available image classification datasets show that the proposed method achieves a transfer gain of 3.3%compared to the best baseline method on single-source transfer tasks on different datasets.Besides,the proposed method achieves a transfer gain of 2.7%on the multi-source transfer task compared to the best baseline method.Next,for the problem of insufficient model generalization capacity in multi-source transfer learning,this paper proposes a transferable graph knowledge-based multi-source transfer learning method.It first applies the MMD distance to select the source network that is more similar to the target network from multiple source networks to improve the transferability of source networks.Secondly,it proposes to construct network-invariant common subgraph bases based on the common subgraph structures and corresponding label information.Then based on the K-nearest neighbor algorithm,it assigns pseudolabels to the node features of the target network in the node attribute feature space to realize the dual adaptation of attribute and structural features of nodes,which reduces the misclassification of nodes in the target network.Extensive experiments on real-world social networks and citation networks show that the average classification accuracy on 12 multi-source cross-network node classification tasks achieves a 3.22%performance improvement over the best baseline methods.Last,for the problem of weak feature discriminability in multi-source transfer learning,this thesis proposes a data selection-based multi-source transfer learning method.It selects the most transferable source network according to the similarity of node attribute features in the source and the target network.In order to enhance the discriminability of node features,we propose to use node attribute feature alignment to further reduce the distribution difference between the source and target network so that the node features are distributed near the center of the corresponding category or far away from the cluster edge to avoid misclassification.In addition,we propose a parameter transfer strategy to share the model trained in the source network to the target network,which realizes the efficient training of the prediction model in the target network.Extensive experiments on social networks and citation networks show that compared to the best baseline methods,the inferior-and best-performing methods of the proposed framework achieve 3.85%and 4.00%transfer gain on 36 multi-source cross-network node classification tasks,respectively. |
