Classic Federated Learning （FL） algorithms are difficult to achieve good results in scenarios where data is highly heterogeneous. In Personalized FL （PFL）， a new solution was proposed aiming at the problem of data heterogeneity in federated learning， which is to “tailor” a dedicated model for each client. In this way， the models had good performance. However， it brought the difficulty in extending federated learning to new clients at the same time. Focusing on the challenges of performance and scalability in PFL， FedDual， a FL model with dual stream neural network structure， was proposed. By adding an encoder for analyzing the personalized characteristics of clients， this model was not only able to have the performance of personalized models， but also able to be extended to new clients easily. Experimental results show that compared to the classic Federated Averaging （FedAvg） algorithm on datasets such as MNIST and FashionMNIST， FedDual obviously improves the accuracy； on CIFAR10 dataset， FedDual improves the accuracy by more than 10 percentage points， FedDual achieves “plug and play” for new clients without decrease of the accuracy， solving the problem of difficult scalability for new clients.

Time slicing methods in dynamic networks greatly influence the accuracy of community evolution analysis results. As communities vary nonlinearly with time and network topology， both the existing uniform time slicing method and network topology variance-based nonuniform time slicing method are unsatisfactory in capturing community evolution events. Therefore， a nonuniform time slicing method based on prediction of community variance was proposed， where the community variance is quantitatively described by the difference between the community modularity expected to be achieved by the updated network and the community modularity obtained by directly applying the community detection results of the network before changing. Firstly， the prediction model of community modularity was established on the basis of time series analysis. Secondly， with the established model， the expected community modularity of the updated network was predicted， and the prediction value of community variance was obtained. Finally， once the prediction value surpassed a previously set threshold， a new time slice was generated. Experimental results on two real network datasets show that compared with the traditional uniform time slicing method and the nonuniform time slicing method based on network topology variance， on the dynamic network dataset Arxiv HEP-PH， the proposed method identifies community disappearance events 1.10 days and 1.30 days earlier， respectively， and identifies the community forming events 8.34 days and 3.34 days earlier， respectively， and the total number of identified community shrinking and growing events increased by 10 and 1 respectively. On Sx?MathOverflow dataset， the proposed method identifies community disappearance events 3.30 days and 1.80 days earlier， and identifies the community forming events 6.41 days and 2.97 days earlier respectively， and the total number of identified community shrinking and growing events increased by 15 and 7， respectively.

The recognition of spam is one of the main tasks in natural language processing. The traditional methods are based on text features or word frequency， which recognition accuracies mainly depend on the presence or absence of specific keywords. When there are no keywords or errors in recognizing keywords in the spam， the traditional methods have poor recognition performance. Neural network-based methods were proposed. Recognition training and testing were conducted on complex spam. The spams that cannot be recognized by traditional methods were collected and the same amount of normal information was randomly selected from spam messages， advertisement and spam email datasets to form three new datasets without duplicate data. Three models were proposed based on convolutional neural network and recurrent neural network and tested on three new datasets for spam recognition. The experimental results show that the neural network-based models learned better semantic features from the text and achieved the accuracies of more than 98% on all three datasets， which are significantly higher than those of the traditional methods， such as Naive Bayes （NB）， Random Forest （RF） and Support Vector Machine （SVM）. The experimental results also show that different neural networks are suitable for text classification with different lengths. The models composed of recurrent neural networks are good at recognizing text with sentence length， the models composed of convolutional neural networks are good at recognizing text with paragraph length， and the models composed of both neural networks are good at recognizing text with chapter length.

Aiming at the problem that the steganalysis of images is difficult, and the existing detection models are difficult to make a targeted analysis of steganography regions of images, a method for image steganalysis based on saliency detection was proposed. In the proposed method, the saliency detection was used to guide the steganalysis model to focus on the image features of steganography regions. Firstly, the saliency detection module was used to generate saliency regions of the image. Secondly, the region filter module was used to filter the saliency images with a high degree of coincidence with the steganography regions, and image fusion technology was used to fuse them with the original images. Finally, the quality of training set was improved by replacing the error detection images with their corresponding saliency fusion images, so as to improve the training effect and detection ability of the model. The experiments were carried out on BOSSbase1.01 dataset. The dataset was embedded by image adaptive steganography algorithms in spatial domain and JPEG domain respectively, and experimental results show that the proposed method can effectively improve the the detection accuracy for deep learning-based steganalysis model by 3 percentage points at most. The mismatch test was also carried out on IStego100K dataset to further verify the generalization ability of the model and improve its application value. According to the result of the mismatch test, the proposed method has certain generalization ability.

Aiming at how to merge the nodes' attributes and network structure information to realize the classification of social network nodes, a social network node classification algorithm based on graph encoder network was proposed. Firstly, the information of each node was propagated to its neighbors. Secondly, for each node, the possible implicit relationships between itself and its neighbor nodes were mined through neural network, and these relationships were merged together. Finally, the higher-level features of each node were extracted based on the information of the node itself and the relationships with the neighboring nodes and were used as the representation of the node, and the node was classified according to this representation. On the Weibo dataset, compared with DeepWalk model, logistic regression algorithm and the recently proposed graph convolutional network, the proposed algorithm has the classification accuracy greater than 8%; on the DBLP dataset, compared with multilayer perceptron, the classification accuracy of this algorithm is increased by 4.83%, and is increased by 0.91% compared with graph convolutional network.

In view of the poor separability and carrier recovery distortion of the current reversible data hiding technology, a novel scheme of separable reversible data hiding was proposed in encrypted domain. Hexadecimal data was embedded by recoding in the cipher text redundancy by the weight of the encrypted domain and the recoding of the encrypted data in Ring-Learning With Errors (R-LWE) algorithm. With embedded cipher text, the additional data was extracted by using data-hiding key, and the original data was recovered losslessly by using encryption key, and the processes of extraction and decryption were separable. By deducing the error probability of the scheme, the parameters in the scheme which directly related to the scheme's correctness were mainly discussed, and reasonable values of the parameters were got by experiments. The experimental results demonstrate that the proposed scheme can better guarantee the reversibility losslessly and 1 bit plaintext data can maximally load 4 bits additional data in encrypted domain.

Reversible data hiding is a new research direction of information hiding technology. Reversible data hiding in encrypted domain is a significant point which combines the technologies of the signal processing in encrypted domain and information hiding and can play an important role of double insurance for information security in data processing. In particular with the adoption of cloud services, reversible data hiding in encrypted domain has become a focused issue to achieve privacy protection in the cloud environment. Concerning the current technical requirements, the background and the development of reversible data hiding were introduced in encrypted domain, and the current technical difficulties were pointed out and analysed. By studying on typical algorithms of various types, the reversible data hiding algorithms in encrypted domain were systematically classified and their technical frameworks, characteristics and limitations of different applications were analysed. Finally, focused on the technology needs and difficulties, several future directions in this field were proposed.