Aiming at the problems of low hardware resource utilization and high latency of Convolutional Neural Network (CNN) when performing inference on heterogeneous platforms, a self-adaptive partitioning and scheduling method of CNN inference model was proposed. Firstly, the key operators of CNN were extracted by traversing the computational graph to complete the adaptive partition of the model, so as to enhance the flexibility of the scheduling strategy. Then, based on the performance measurement and the critical path-greedy search algorithm, according to the sub-model running characteristics on the CPU-GPU heterogeneous platform, the optimal running load was selected to improve the sub-model inference speed. Finally, the cross-device scheduling mechanism in TVM (Tensor Virtual Machine) was used to configure the dependencies and running loads of sub-models in order to achieve adaptive scheduling of model inference, and reduce the communication delay between devices. Experimental results show that on GPU and CPU, compared to the method optimized by TVM operator, the proposed method improves the inference speed by 5.88% to 19.05% and 45.45% to 311.46% with no loss of model inference accuracy.
Concerning of the problems of high cost and unstable detection results of the traditional malicious code detection methods, a multi-neural network malicious code detection model based on depthwise separable convolution was proposed. By using the Depthwise Separable Convolution (DSC), SENet (Squeeze-and-Excitation Network) channel attention mechanism and Grey Level Co-occurrence Matrix (GLCM), three lightweight neural networks were connected with GLCM in parallel to detect malicious code families and their variants, then the detection results of multiple strong classifiers were fused via Naive Bayes classifier to improve the detection accuracy while reducing the computational cost. Experimental results on the hybrid dataset of MalVis + benign data show that the proposed model achieved the accuracy of 97.43% in the detection of malicious code families and their variants, which was 6.19 and 2.29 percentage points higher than those of ResNet50 and VGGNet models respectively, while its parameter quantity is only 68% of that of ResNet50 model and 13% of that of VGGNet model. On malimg dataset, the detection accuracy of this model achieved 99.31%. In conclusion, the proposed model has good detection effect with reduced parameters.
In massive Multiple-Input Multiple-Output (MIMO) systems, Minimum Mean Square Error (MMSE) detection algorithm has the problems of poor adaptability, high computational complexity and low efficiency on the reconfigurable array structure. Based on the reconfigurable array processor developed by the project team, a parallel mapping method based on MMSE algorithm was proposed. Firstly, a pipeline acceleration scheme which could be highly parallel in time and space was designed based on the relatively simple data dependency of Gram matrix calculation. Secondly, according to the relatively independent characteristic of Gram matrix calculation and matched filter calculation module in MMSE algorithm, a modular parallel mapping scheme was designed. Finally, the mapping scheme was implemented based on Xilinx Virtex-6 development board, and the statistics of its performance were performed. Experimental results show that, the proposed method achieves the acceleration ratio of 2.80, 4.04 and 5.57 in Quadrature Phase Shift Keying (QPSK) uplink with the MIMO scale of 128 × 4 , 128 × 8 and 128 × 16 , respectively, and the reconfigurable array processor reduces the resource consumption by 42.6% compared with the dedicated hardware in the 128 × 16 massive MIMO system.
The traditional image encryption with scrambling-diffusion structure is usually divided into two independent steps of scrambling and diffusion, which are easy to be cracked separately, and the encryption process has weak nonlinearity, resulting in poor security of the algorithm. Therefore, a scrambling diffusion synchronous image encryption algorithm with strong nonlinearity was proposed. Firstly, a new sine-cos chaotic mapping was constructed to broaden the range of control parameters and improve the randomness of sequence distribution. Then, the exclusive-OR sum of plaintext pixels and chaotic sequence was used as the initial chaotic value to generate chaotic sequence, and this chaotic sequence was used to construct the network structures of different pixels of different plaintexts. At the same time, the diffusion value was used to dynamically update the network value to make the network dynamic. Finally, the single pixel serial scrambling-diffusion was used to generate cross-effect between scrambling and diffusion,and the overall synchronization of scrambling and diffusion, so as to effectively resist separation attacks. In addition, the pixel operations were transferred according to the network structure, which made the serial path nonlinear and unpredictable, thereby ensuring the nonlinearity and security of the algorithm. And the adjacent node pixels sum was used to perform dynamic diffusion in order to improve the correlation of the plaintext. Experimental results show that the proposed algorithm has high encryption security, strong plaintext sensitivity, and is particularly effective in anti-statistical attack, anti-differential attack and anti-plaintext attack.
In view of the problems that the traditional general graph matching search is inefficient, and refractive index data cannot be positioned fast in large data environment, a distributed massive molecular retrieval model based on consistent Hash function was established. Combined with the characteristics of molecular storage structures, to improve retrieval efficiency of molecules, the continuous refractive index was discretized by fixed width algorithm to establish high-speed Hash index, and the distributed massive retrieval system was realized. The size of dataset was effectively reduced, and Hash collision was handled according to the visiting frequency. The experimental results show that, in the chemical data containing 200 thousand structures of molecules, the average time of this method is about five percent of the traditional general graph matching search. Besides, the model has the steady performance with high scalability. It is applicable to retrieve high-frequency molecules in accordance with refractive index under the environment of massive data.
For improving the accuracy and validity of social E-commerce recommendation services, a trust-aware collaborative filtering recommendation method was proposed with considering the factors that influence the trust relationship of users in social E-commerce, such as transaction evaluation score, transaction frequency, transaction amount, direct trust and recommended reputation. The belief factor was introduced to compute the trust relationship of social E-commerce users, the cosine similarity method was used to calculate the similarity of the users, the harmonic factor was used to synthesize the influence of the trust relationship and similarity on the users, the Mean Absolute Error (MAE), rating coverage and user coverage were used as the evaluation indexes. The experimental results show that the accuracy of the trust-aware collaborative filtering method is superior to the traditional collaborative filtering method and the regularized matrix factorization based collaborative filtering recommendation method in that the MAE reduces to 0.162, and the rating coverage and user coverage rise to 77% and 80% respectively. This proves that the trust-aware collaborative filtering method can solve the problem of recommending the commodities with less transaction evaluation.
To solve the synchronization problem of audio and video, an improved algorithm of audio-video synchronization coding based on H.264 inter-frame prediction was proposed. The algorithm introduced the concept of variable code length. The audio encoding data was divided into several code groups, and each code group had 2 or 3 bits of embedded data. In the stage of H.264 inter-frame prediction, the mappings between various variable size blocks and the data of code groups were based on formula. The coding method was dynamically determined for the macro block modes coding according to embedded data, and a proposed decoding method could extract the corresponding data according to the mapping relationship. Finally, the 4×4 macro block mode was used to indicate the end of the audio data.The experimental results show that the proposed algorithm enables the Peak Signal-to-Noise Ratio (PSNR) of video samples to reduce by 0.031dB, the bit rate to increase by 5.16% and the overhead to increase by 1.97%, but the embedded audio data can be correctly and completely extracted. Therefore,the algorithm can implement the synchronization of audio and video coding while increasing the data embedding capacity, maintaining the quality of video, ensuring the correctness and completeness of the data.