The free spread of public opinions may lead to the occurrence of cyber collective behaviors， which are easy to cause negative social impacts and threaten public security. Therefore， the establishment of network public opinion monitoring and early warning mechanism is necessary to prevent and control the spread of public opinions and maintain social stability. Firstly， by analyzing the formation mechanism of rumors， a prediction index system of public opinion development was constructed. Secondly， the multifactor GM（1， n ） model was established to predict the development trend of the public opinion. Then， the prediction model was improved by combining with metabolism theory and Markov theory. Finally， using the “Xinjiang cotton” event and “Chengdu No.49 middle school” event in Weibo as examples， the abilities of the GM（1， n ） model， the Markov GM（1， n ） model and the metabolic Markov GM（1， n ） model to predict the development of public opinions were compared，and the metabolic Markov GM（1， n ） model was also compared with the random forest model.Experimental results show that the average prediction accuracy of the metabolic Markov GM（1， n ） model is increased by 10.6% and 5.8% compared with those of the original GM（1， n ） model and random forest model respectively. It can be seen that the metabolic Markov GM（1， n ） model has good performance in predicting the development trend of network public opinions.

Most of the existing expert recommendation processes rely on manual matching, which leads to the low accuracy of expert recommendation due to that they cannot fully capture the semantic association between the subject of the project and the interests of experts. To solve this problem, a matching method for academic expertise of project peer review experts was proposed. In the method, an academic network was constructed to establish the academic entity connection, and a meta-path was designed to capture the semantic association between different nodes in the academic network. By using the random walk strategy, the node sequence of co-occurrence association between the subject of the project and the expert research interests was obtained. And through the network representation learning model training, the vector representation with semantic association of the project subject and expert research interests was obtained. On this basis, the semantic similarity was calculated layer by layer according to the hierarchical structure of project subject tree to realize multi-granularity peer review academic expertise matching. Experimental results on the crawled datasets of HowNet and Wanfang papers, an expert review dataset and Baidu Baike word vector dataset show that this method can enhance the semantic association between the project subject and expert research interests, and can be effectively applied to the academic expertise matching of project review experts.

Numerous logs generated during the daily deployment and operation process in cloud computing platforms help system administrators perform anomaly detection. Common anomalies in cloud workflow include pathway anomalies and time delay anomalies. Traditional anomaly detection methods train the learning models corresponding to the two kinds of anomaly detection tasks respectively and ignore the correlation between these two tasks, which leads to the decline of the accuracy of anomaly detection. In order to solve the problems, an anomaly detection method based on multi-task temporal convolutional network was proposed. Firstly, the event sequence and time sequence were generated based on the event templates of log stream. Then, the deep learning model based on the multi-task temporal convolutional network was trained. In the model, the event and the time characteristics were learnt in parallel from the normal system execution processes by sharing the shallow layers of the temporal convolutional network. Finally, the anomalies in the cloud computing workflow were analyzed, and the related anomaly detection logic was designed. Experimental results on the OpenStack dataset demonstrate that, the proposed method improves the anomaly detection accuracy at least by 7.7 percentage points compared to the state-of-art log anomaly detection algorithm DeepLog and the method based on Principal Component Analysis (PCA).

Focusing on the issue of low efficiency of SKINNY encryption algorithm in Central Processing Unit(CPU), a fast implementation method was proposed based on Graphic Processing Unit(GPU). In the first place, an optimization scheme was proposed by combining the structural characteristics of SKINNY algorithm, and one whole calculation, where the whole calculation was integrated by 5 step-by-step operations. Moreover, the characteristics of the Electronic CodeBook(ECB) mode and counter(CTR) mode of this algorithm were analyzed, and the parallel design schemes such as parallel granularity and memory allocation were given. Experimental results illustrate that the efficiency and throughput of SKINNY algorithm implemented by Computing Unified Device Architecture(CUDA) are significantly improved, when compared to the algorithm with the traditional CPU implementation. More specifically, for data size of 16 MB or large size, the SKINNY algorithm implementation with ECB mode achieves maximum efficiency improvement of 99.85% and maximum speedup ratio of 671. On the other hand, the SKINNY algorithm implementation with CTR mode achieves maximum efficiency improvement of 99.87% and maximum speedup ratio of 765. In particular, the throughput of the proposed SKINNY-256(ECB) parallel algorithm has 1.29 times and 2.55 times of those of the existing AES-256(ECB) and SKINNY_ECB parallel algorithms, respectively.

In Fixed Word-Aligned (FWA) inverted index compression algorithms of Web search engines, due to the "greedy" block partition strategy of the inverted list and the interleaved storage of the codeword information, it is difficult for the algorithm to achieve the optimal compression performance. To solve the above problem, an FWA partition compression algorithm based on Directed Acyclic Graph (DAG) was proposed. Firstly, considering the small integer information in the inverted list brought by the clustering characteristics of Web pages, a novel FWA compression format with data area of 64-bit blocks was designed. In this compression format, the data area was divided into 16 storage modes suitable for continuous small integer compression through 4-bit selector area, and the selector area and data area in each block of the inverted list were stored separately, so as to ensure good batch decompression performance. Secondly, a DAG described Word-Aligned Partition (WAP) algorithm was proposed on the basis of the new compression format. In the algorithm, the inverted list block partitioning problem was regarded as a Single-Source Shortest-Path (SSSP) problem by DAG, and by considering the constraints of various storage modes of data area in FWA compression format, the structure and recursive definition of the SSSP problem were determined. Thirdly, the dynamic programming technique was used to solve the problem of SSSP and generate the pseudo-code and algorithm complexity of the optimal partition vector. The original storage modes of traditional FWA algorithms such as S9, S16 and S8b were partitioned and optimized based on DAG, and the computational complexities of the algorithms before and after optimization were compared and analyzed. Finally, the compression performance experiments were conducted with simulation integer sequence data and Text REtrieval Conference (TREC) GOV2 Web page index data. Experimental results show that, compared with the traditional FWA algorithms, the DAG based FWA partition algorithm can improve the compression ratio and decompression speed by batch decompression and partition optimization technology. At the same time, it can obtain a higher compression ratio than the traditional Frame Of Reference (FOR) algorithms for the compression of continuous small integer sequence.

Focused on the low analysis efficiency of Correlation Power Analysis (CPA) interfered by noise, a stepwise CPA scheme was proposed. Firstly, the utilization of information in CPA was improved by constructing a new stepwise scheme. Secondly, the problem that the accuracies of previous analyses were not guaranteed was solved by introducing the confidence index to improve the accuracy of each analysis. Finally, a stepwise CPA scheme was proposed based on the structure of SM4 cryptographic algorithm. The results of simulation experiments show that, on the premise of the success rate up to 90%, stepwise CPA reduces the demand of power traces by 25% compared to classic CPA. Field Programmable Gate Array (FPGA) based experiments indicate that the ability of stepwise CPA to recover the whole round key is very close to the limit of expanding the search space to the maximum. Stepwise CPA can reduce the interference of noise and improve the efficiency of analysis with a small amount of calculation.

Aiming at the problem of the present mainstream adversarial attack algorithm that the attack invisibility is reduced by disturbing the global image features, an untargeted attack algorithm named PS-MIFGSM (Perceptual-Sensitive Momentum Iterative Fast Gradient Sign Method) was proposed. Firstly, the areas of the image focused by Convolutional Neural Network (CNN) in the classification task were captured by using Grad-CAM algorithm. Then, MI-FGSM (Momentum Iterative Fast Gradient Sign Method) was used to attack the classification network to generate the adversarial disturbance, and the disturbance was applied to the focus areas of the image with the non-focus areas of the image unchanged, thereby, a new adversarial sample was generated. In the experiment, based on three image classification models Inception_v1, Resnet_v1 and Vgg_16, the effects of PS-MIFGSM and MI-FGSM on single model attack and set model attack were compared. The results show that PS-MIFGSM can effectively reduce the difference between the real sample and the adversarial sample with the attack success rate unchanged.

For the fact that information groups arrive at the system in a continuous time, a two-level polling service model with different priorities was proposed for the business problems of different priorities in the polling system. Firstly, gated service was used in sites with low priority, and exhaustive service was used in sites with high priority. Then, when high priority turned into low priority, the transmission service and the transfer query were processed in parallel to reduce the time cost of server during query conversion, improving the efficiency of polling system. Finally, the mathematical model of system was established by using Markov chain and probabilistic parent function. By accurately analyzing the mathematical model, the expressions of average queue length and average waiting time of each station of continuous-time two-level service system were obtained. The simulation results show that the theoretical calculation value was approximately equal to the experimental simulation value, indicating that the theoretical analysis is correct and reasonable. The model provides high-quality services for high-priority sites while maintaining the quality of services in low-priority sites.

In order to enhance the network representation learning quality with node profile information, and focus on the problems of semantic information dispersion and incompleteness of node profile attribute information in social network, a network representation learning algorithm incorporated with node profile information was proposed, namely NPA-NRL. Firstly, attribute information were encoded by one-hot encoding, and a data augmentation method of random perturbation was introduced to overcome the incompleteness of node profile attribute information. Then, attribute coding and structure coding were combined as the input of deep neural network to realize mutual complementation of the two types of information. Finally, an attribute similarity measure function based on network homogeneity and a structural similarity measure function based on SkipGram model were designed to mine fused semantic information through joint training. The experimental results on three real network datasets including GPLUS, OKLAHOMA and UNC demonstrate that, compared with the classic DeepWalk, Text-Associated DeepWalk (TADW), User Profile Preserving Social Network Embedding (UPP-SNE) and Social Network Embedding (SNE) algorithms, the proposed NPA-NRL algorithm has a 2.75% improvement in average Area Under Curve of ROC (AUC) value on link prediction task, and a 7.10% improvement in average F1 value on node classification task.

Concerning the problem that a predefined time window is usually used to mine simple association dependencies between events in the traditional frequent episode discovery, which cannot effectively handle interleaved temporal correlations between events, a concept of time-lag episode discovery was proposed. And on the basis of frequent episode discovery, Adjacent Event Matching set (AEM) based time-lag episode discovery algorithm was proposed. Firstly, a probability statistical model introduced with time-lag was introduced to realize event sequence matching and handle optional interleaved associations without a predefined time window. Then the discovery of time lag was formulated as an optimization problem which can be solved iteratively to obtain time interval distribution between time-lag episodes. Finally, the hypothesis test was used to distinguish serial and parallel time-lag episodes. The experimental results show that compared with Iterative Closest Event (ICE) algorithm which is the latest method of time-lag mining, the Kullback-Leibler (KL) divergence between true and experimental distributions discovered by AEM is 0.056 on average, which is decreased by 20.68%. AEM algorithm measures the possibility of multiple matches of events through a probability statistical model of time lag and obtains a one-to-many adjacent event matching set, which is more effective than the one-to-one matching set in ICE for simulating the actual situation.

The composition of neighborhoods is crucial for the spatial domain-based Graph Convolutional Network (GCN) model. To solve the problem that the structural influence is not considered in the neighborhood ordering of nodes in the model, a novel neighborhood selection strategy was proposed to obtain an improved GCN model. Firstly, the structurally important neighborhoods were collected for each node and the core neighborhoods were selected hierarchically. Secondly, the features of the nodes and their core neighborhoods were organized into a matrix. Finally, the matrix was sent to deep Convolutional Neural Network (CNN) for semi-supervised learning. The experimental results on Cora, Citeseer and Pubmed citation network datasets show that, the proposed model has a better accuracy in node classification tasks than the model based on classical graph embedding and four state-of-the-art GCN models. As a spatial domain-based GCN, the proposed model can be effectively applied to the learning tasks of large-scale networks.

Traditional network operation and maintenance evaluation method has two problems. First, it is too dependent on domain experts' experience in indicator selection and weight assignment, so that it is difficult to obtain accurate and comprehensive assessment results. Second, the network operation and maintenance quality involves data from multiple manufacturers or multiple devices in different formats and types, and a surge of users brings huge amounts of data. To solve the problems mentioned above, an indicator selection method based on correlation was proposed. The method focuses on the steps of indicator selection in the process of evaluation. By comparing the strength of the correlation between the data series of indicators, the original indicators could be classified into different clusters, and then the key indicators in each cluster could be selected to construct a key indicators system. The data processing methods and weight determination methods without human participation were also utilized into the network operation and maintenance quality evaluation model. In the experiments, the indicators selected by the proposed method cover 72.2% of the artificial indicators. The information overlap rate is 31% less than the manual indicators'. The proposed method can effectively reduce human involvement, and has a higher prediction accuracy for the alarm.

Focusing on the linear decomposition of the S-box layer in Keccak algorithm, a new linear structure construction method was proposed based on the algebraic properties of the S-box. Firstly, to ensure the state data was still linear with that after this linear structure, some constraints about input bits of S-box needed to be fixed. Then, as an application of this technique, some new zero-sum distinguishers of round-reduced Keccak were constructed by combining the idea of meet-in-the-middle attack. The results show that a new 15-round distinguisher of Keccak is found, which extends 1-round forward and 1-round backward. This work is consistent with the best known ones and its complexity is reduced to 2 ²⁵⁷. The new distinguisher, which extends 1-round forward and 2-round backward, has the advantages of more free variables and richer distinging attack combinations.

Manual rumor detection often consumes a lot of manpower and material resources, and there will be a long detection delay. At present, the existing rumor detection models construct features manually according to the content, user attributes, and pattern of the rumor transmission, which can not avoid one-sided consideration, waste of human and other phenomena. To solve this problem, a rumor detection model based on Convolutional Neural Network (CNN) was presented. The rumor events in microblog were vectorized. The deep features of text were mined through the learning and training in hidden layer of CNN to avoid the problem of feature construction, and those features that were not easily found could be found to produce better results. The experimental results show that the proposed method can accurately identify rumor events, and it is better than Support Vector Machine (SVM), Recurrent Neural Network (RNN) and other contrast algorithms in accuracy rate, precision rate and F1 score.

According to the problem of premature convergence and local optimum in Firefly Algorithm (FA), this paper came up with a kind of multi-group firefly algorithm based on simulated annealing mechanism (MFA_SA), which equally divided firefly populations into many child populations with different parameter. To prevent algorithm fall into local optimum, simulated annealing mechanism was adopted to accept good solutions by the big probability, and keep bad solutions by the small probability. Meanwhile, variable distance weight was led into the process of population optimization to dynamically adjust the "vision" of firefly individual. Experiments were conducted on 5 kinds of benchmark functions between MFA_SA and three comparison algorithms. The experimental results show that, MFA_SA can find the global optimal solutions in 4 testing function, and achieve much better optimal solution, average and variance than other comparison algorithms. which demonstrates the effectiveness of the new algorithm.

Since the Multi-Label K Nearest Neighbor (ML-KNN) classification algorithm ignores the correlation between labels, a multi-label classification algorithm by exploiting label correlation named CML-KNN was proposed. Firstly, the conditional probability between each pair of labels was calculated. Secondly, the conditional probabilities of predicted labels and the conditional probability of the label to be predicted were ranked, then the maximum was got. Finally, a new classification model by combining Maximum A Posteriori (MAP) and the product of the maximum and its corresponding label value was proposed and the new label value was predicted. The experimental results show that the performance of CML-KNN on Emotions dataset outperforms the other four algorithms, namely ML-KNN, AdaboostMH, RAkEL, BPMLL, while only two evaluation metric values are lower than those of ML-KNN and RAkEL on Yeast and Enron datasets. The experimental analyses show that CML-KNN obtains better classification results.

Feature point matching is of central importance in feature-based image registration algorithms such as Scale-Invariant Feature Transform (SIFT) algorithm. Since most of the existed feature matching algorithms are not so powerful and efficient in mismatch removing, in this paper, a mismatch removal algorithm was proposed which adopted the depth information in an image to improve the performance. In the proposed approach, the depth map of an acquired image was produced using the clues of defocusing blurring effect, and machine learning algorithm, followed by SIFT feature point extraction. Then, the correct feature correspondences and the transformation between two feature sets were iteratively estimated using the RANdom SAmple Consensus (RANSAC) algorithm and exploiting the rule of local depth continuity. The experimental results demonstrate that the proposed algorithm outperforms conventional ones in mismatch removing.

Due to the large requirement for memory and the high complexity of computation, JPEG2000 can not be used in many conditions. The line-based wavelet transform was proposed and accepted because it required lower memory without affecting the result of wavelet transform. In this paper, the improved lifting scheme was used to perform wavelet transform to replace Mallat algorithm in the original linebased wavelet transform, the corresponding context-based arithmetic coding was discussed here too. As a result, considerable reduction of memory and computational costs can be achieved.

Due to the high complexity of computation, JPEG2000 cannot be used in many conditions. As an important part of JPEG2000, EBCOT algorithm performs the scanning 3 times in each bit plane and codes all coefficients in 3 coding passes, wasting lots of operation time. A scanning algorithm was introduced and all the coefficients of one bit plane could be coded in one scanning. Firstly, to deal with the problems caused by this algorithm, two significance state variables rather than a single variable were introduced, Secondly, the "stripe-causal" mode was used to form the context. Finally the states of MQ coder and context were stored separately according to 3 coding passes. Thus one scanning method could improve the coding efficiency greatly while keeping the excellent performance of JPEG2000.