To address the problems of data sparsity and cold start in traditional Collaborative Filtering （CF）， as well as errors caused by various transformations in the process of generating result matrices using matrix factorization methods， a Low-rank and Sparse Matrix Factorization （LSMF） paper recommendation method with mixed information enhancement was proposed. Firstly， pre-trained document-level representation learning and citation aware converter — SPECTER （Scientific Paper Embeddings using Citation-informed TransformERs） was used to learn the representation of papers， and then the similarity matrix among papers was calculated and constructed. Secondly， the similarity matrix and citation matrix were added together to form a mixed information matrix， and then the content similarity information and citation information were integrated into the paper-author matrix through matrix multiplication. Finally， the recommendation list was obtained by using LSMF model to decompose the paper-author matrix. Experimental results on ACL Anthology Network （AAN） and DBLP datasets show that the proposed method achieves better recommendation performance， and the way of introducing content information and citation information in the proposed method can be equally applicable to other matrix factorization models. For Non-negative Matrix Factorization （NMF）， Singular Value Decomposition （SVD）， Low-rank and Sparse Matrix Completion （LSMC）， and Go Decomposition （GoDec）， the Recall values of the top 30 recommended results （R@30） of these models with mixed information are increased by 18.72，7.43，11.53，14.62 and 20.58， 2.11， 7.91， 5.01 percentage points， respectively， compared with those of the original models on the two datasets.

On the basis of the Fast Sparse Coding (FSC) model, considering the maximum sparse distribution of feature coefficients and the orthogonality of feature bases of an image, a Modified FSC (MFSC) model was proposed in this paper. This FSC algorithm was based on iteratively solving two convex optimization problems: L ₁-norm based regularized least square problem and L ₂-norm based constrained least square problem, and it can realize the learning of complete bases and overcomplete bases, as well as efficiently extract the features of images. Moreover, the convergence speed of FSC is quicker than that of Basic Sparse Coding (BSC). The images of natural scene and palmprint were used to test the property of FSC algorithm proposed by the authors in feature extraction, and then the extracted features were utilized to image reconstruction. Compared with reconstructed images obtained by BSC, the experimental results verify the validity of the modified FSC in quickly extracting image features.

To ensure the security and robustness of hiding image information under the conditions of visibility, a digital image hiding method based on a chaotic system with variable parameters was proposed. Firstly, with the pseudo-random sequences generated by the chaos system, a hiding image was encrypted. Secondly, the one-dimensional vector which transformed from the template image was divided into a number of equal parts according to the size of the hiding image, and each part corresponding to each pixel of the hiding image. Finally, with the new chaotic sequences and digital image blending technique, the pixel of hiding image was blended into a position of the part. This digital image hiding technique improves the imperceptibility and security of hiding image. Simulation results show that the proposed technique has good imperceptibility and high robustness.

In anomaly detection, utilizing support vector machines can make detection system have good generalization ability in situation of small sample. But appropriate parameters are very crucial to the learning results and generalization ability of support vector machines. And many irrelevant and redundant features degrade the performance of classification. Thus an approach that applied immune algorithm to optimize parameters of SVM(Support Vector Machine) and feature selection was proposed. Immune algorithm is an efficient random global optimization technique. It has nice performances such as avoiding local optimum, high precision solution, and quick convergence. The simulation results show that immune algorithm can improve the detection accuracy and meanwhile shorten the testing time.