A new network architecture of SCN (service customized networking) was systematically expounded, which provided a novel network underlying capability and usage method for Internet applications.TCP/IP network architecture served the application with “best effort” quality of service, while SCN converted the subject-object relationship between the application and network that allowed the application to “on-demand customize” the QoS of networks.From the perspective of applications, three major connotations “declarable”, “fine-grained” and “end-to-end” were excavated to deduce the overall design of SCN, where a concrete SCN architecture and a feasible SCN system realization were present.In future, SCN can be applied in scenarios of man-machine-things, such as remote industrial control, and augment reality, to provide a novel and ideal means for NaaS (network as a service).

In uplink non-orthogonal multiple access (NOMA)-based unmanned aerial vehicle (UAV)-assisted mobile edge computing (MEC) system, the successive interference cancellation (SIC) order of NOMA became a bottleneck limiting the transmission performance of task offload in uplink link.To reduce the energy consumption of the system, the SIC order was discussed and the optimal SIC order based on channel gain and task delay constraint was proposed.The optimization problem of minimizing the system energy consumption was proposed based on the optimal SIC order while satisfying the constraints of the given task delay of the device, the maximum transmit power constraint of the device, and the UAV trajectory.Since the problem was a complex non-convex problem, an alternating optimization method was adopted to solve the optimization problem to achieve power allocation and UAV trajectory optimization.A low-complexity algorithm based on matching theory was proposed to obtain the optimal device grouping in different time slots.Simulation results show that the optimal SIC order can realize smaller system energy consumption under the same task delay constraint compared with other SIC order, the proposed low-complexity device grouping algorithm can obtain the optimal device grouping.

To solve the privacy-preserving problem of both the user and the server in indoor positioning, outsourcing part of the calculation to cloud server in the process of using Paillier encryption was considered.The scheme not only protected the privacy of the user and the positioning server, but also avoided excessive computing and communication overhead.The main idea of the scheme was that the fingerprint database in the offline stage was established by the server firstly.The k-anonymity algorithm was combined with Paillier encryption in the online stage by the user, and the encrypted Wi-Fi fingerprints were sent to the positioning server.An aggregation of the received Wi-Fi fingerprints and database fingerprints were performed by the server.Then they were outsourced to the cloud server for decryption and distance calculation by the positioning server.Finally, the positioning result was obtained.Theoretical analysis and experimental results show that the proposed scheme is safe, effective and practical.

To solve the problems of the privacy information leakage caused by the deliberate or inadvertent retention of information when information was frequently exchanged across nodes and systems in a ubiquitous sharing environment, a cross-domain multi-copy of flow discovery mechanism based on dual certificate storage was proposed, which could trace the propagation path and channel, and construct a multi-copy propagation graph of the information.Depending on the timing and method of certification, the dual certification was comprised active circulation certification and passive operation certification.Before the information was shared, the information sharer actively recorded the propagation path and method to generate active circulation certification records.Before the information was operated, the system automatically recorded the propagation path to generate passive operational certification records.Compared with single certificate storage, the dual certificate storage could improve the integrity and authenticity of the constructed multi-copy propagation graph of information, and could detect nodes with abnormal certificate storage behavior and provide disposals.Based on the theory of social punishment, the effectiveness of abnormal certificate storage behavior detection and handling was demonstrated.A prototype system for multi-copy discovery of OFD with dual certificate storage is developed, the improvement of information dissemination graph construction integrity by the proposed mechanism is verified.

Considering the problems of traditional intrusion detection methods limited by the class imbalance of datasets and the poor representation of selected features, a detection method based on VAE-CWGAN and fusion of statistical importance of features was proposed.Firstly, data preprocessing was conducted to enhance data quality.Secondly, a VAE-CWGAN model was constructed to generate new samples, addressing the problem of imbalanced datasets, ensuring that the classification model no longer biased towards the majority class.Next, standard deviation, difference of median and mean were used to rank the features and fusion their statistical importance for feature selection, aiming to obtain more representative features, which made the model can better learn data information.Finally, the mixed data set after feature selection was classified through a one-dimensional convolutional neural network.Experimental results show that the proposed method demonstrates good performance advantages on three datasets, namely NSL-KDD, UNSW-NB15, and CIC-IDS-2017.The accuracy rates are 98.95%, 96.24%, and 99.92%, respectively, effectively improving the performance of intrusion detection.

Aiming at the problem of low detection accuracy or missing detection caused by dense vehicles and small scale of distant vehicles in the visual detection of urban road scenes, an improved YOLOv3-SPP algorithm was proposed to optimize the activation function and take DIOU-NMS Loss as the boundary frame loss function to enhance the expression ability of the network.In order to improve the feature extraction ability of the proposed algorithm for small targets and occluding targets, the void convolution module was introduced to increase the receptive field of the target.Based on the experimental results, the proposed algorithm improves the mAP by 1.79% when detecting vehicle targets, and also effectively reduce the missing phenomenon when detecting tight vehicle targets.

Due to software and hardware differentiation, the problem of false positives mistakenly identified as network attack behavior caused by inconsistent mimic decision results frequently occurs.Therefore, a mimic decision method based on deep learning was proposed.By constructing an unsupervised autoencoder-decoder deep learning model, the deep semantic features of diverse normal response data were explored from different executions and its statistical rules were analyzed and summarized.Additionally, the offline learning-online decision-making mechanism and the feedback optimization mechanism were designed to solve false positive problem, thereby accurately detecting network attacks and improving target system security resilience.Since statistical rules of normal response data was understood and mastered by deep learning model, the mimic decision results among different executions could remain consistent, indicating that the target system was in a secure state.However, once the target system was subjected to a network attacks, the response data outputted by the different executions was deviated from statistical distribution of deep learning model.Therefore, inconsistent mimic decision results were presented, indicating that the affected execution was under attack and the target system was exposed to potential security threats.The experiments show that the performance of the proposed method is significantly superior to the popular mimic decision methods, and the average prediction accuracy is improved by 14.89%, which is conducive to integrating the method into the mimic transformation of real application to enhance the system’s defensive capability.

To solve the problem of the weak spectrum-cognitive ability caused by monitoring angle, direction resolution, limited processing ability and peak power for a low-earth-orbit (LEO) satellite, a multi-satellite cooperative spectrum cognitive method integrating Stackelberg game and federated learning was proposed.Firstly, considering the available computing resource, cognitive performance, processing and transmission delay of each spectrum cognitive satellite, a cooperative-satellite selection and computing-resource allocation algorithm was built for multiple spectrum-cognitive tasks.Secondly, based on the selected satellites and the allocated computing resources, a low-complexity multi-satellite cooperative spectrum cognitive strategy was further designed, which could automatically sense the spectrum holes, and detect interference as well as identify the modulation mode.Simulation results demonstrate that compared to the single-node cognitive method, the designed multi-satellite cooperative spectrum cognitive strategy can obtain a better cognitive performance.Moreover, comparing with the existing model, the model utilized in the designed strategy can effectively achieve 96.69% and 93.32% lower number of parameters and required floating point operations per second, whilst maintaining the performance.

To solve the problem that advanced persistent threat (APT) in the Internet of things (IoT) environment had the characteristics of strong concealment, long duration, and fast update iterations, it was difficult for traditional passive detection models to quickly search, a hypergraph Transformer threat-hunting network (HTTN) was proposed.The HTTN model had the function of quickly locating and discovering APT attack traces in IoT systems with long time spans and complicated information concealment.The input cyber threat intelligence (CTI) log graph and IoT system kernel audit log graph were encoded into hypergraphs by the model, and the global information and node features of the log graph were calculated through the hypergraph neural network (HGNN) layer, and then they were extracted for hyperedge position features by the Transformer encoder, and finally the similarity score was calculated by the hyperedge, thus the threat-hunting of APT was realized in the network environment of the Internet of things system.It is shown by the experimental results in the simulation environment of the Internet of things that the mean square error is reduced by about 20% compared to mainstream graph matching neural networks, the Spearman level correlation coefficient is improved by about 0.8%, and improved precision@10 is improved by about 1.2% by the proposed HTTN model.

To solve the problems of energy limitation, spectrum shortage, and serious co-frequency interference in the unmanned aerial vehicle (UAV)-assisted D2D (device-to-device) communication network under passive eavesdropping, the concept of the network secrecy energy efficiency was introduced by taking into account the secrecy rate and transmit power constraints of each user.Based on the physical layer secure transmission theory, the joint optimization problem of spectrum multiplexing and power control strategies was investigated to maximize the network secrecy energy efficiency.The problem was a nonlinear nonconvex problem, a resource allocation algorithm combining Dinkelbach’s algorithm and the convex approximation algorithm was proposed to transform the nonconvex optimization problem into a geometric planning problem through the convex approximation algorithm and ensure that the original problem converges to the global optimal solution by Dinkelbach’s algorithm.Simulation results show that the proposed algorithm can effectively mitigate the co-channel interference, improve the network secrecy energy efficiency, and balance the relationship between the secure communication performance and power consumption of the network.

In the face of the high power consumption issue caused by the dense deployment of base stations in cell-free networks, a wideband cell-free network system with sub-connected active reconfigurable intelligent surface (RIS) was proposed.Firstly, based on the constraints of maximum power consumption at the active RIS, amplification factor, and maximum power consumption at the base station, a joint precoding design problem for the base station and RIS was formulated to maximize the energy efficiency.To solve the non-convex problem, advanced techniques including alternating optimization, block coordinate descent, Lagrange dual reconstruction, and multidimensional complex quadratic transformation were applied to transform the original problem into multiple sub-problems.By iteratively solving each sub-problem, the solutions of the original problem was ultimately obtained.The simulation results validate the effectiveness of the proposed scheme.

Existing routing protocols without awareness of the topology causes excessive retransmissions, energy holes, and long delay, data routing performance was seriously deteriorated.Considering the relation of topology and routing, an intelligent route planning with jointing topology control (IRPJTC) method was proposed.IRPJTC consisted of two part, the virtual force-based adaptive topology control (VFATC), and the PPO-based geographic routing protocol (PPO-GRP).Based on neighbor’s mobility information, the distance between UAVs was adaptively adjusted by VFATC to provide stable links between UAVs.Combined with link stability metric in VFATC, end-to-end delay and energy consumption, a multi-objective reward function was designed by PPO-GRP to train optimal routing strategy.According to the performance study, the proposed IRPJTC reduces existing routing protocols by 12.11% of end-to-end delay, and 4.56% of energy consumption, and has a better energy balance ability.

Those homomorphic encryption schemes supporting single instruction multiple data (SIMD) operations effectively enhance the amortized efficiency of ciphertext computations, yet the structure of ciphertexts leads to high complexity in matrix operations.In many applications, employing plaintext-ciphertext matrix operations can achieve privacy-preserving computing.Based on this, a plaintext-ciphertext matrix multiplication scheme for matrices of arbitrary dimension was proposed.The resulting ciphertext was computed through steps such as encoding the plaintext matrix, transforming the dimensions of the encrypted matrix, etc.Compared to the best-known encrypted matrix multiplication algorithm for square matrices proposed by Jiang et al., the proposed scheme supported matrix multiplication of arbitrary dimension, and consecutive matrix multiplications.Both theoretical analysis and experimental results show that the proposed scheme requires less rotations on ciphertexts and hence features higher efficiency.When applied to a privacy-preserving Bayesian classifier, the proposed scheme can complete classification tasks with higher security parameters and reduced running time.

In view of the problem that it was difficult to separate speaker personality characteristics from semantic content information in any-to-any voice conversion under non-parallel corpus, which led to unsatisfied performance, a voice conversion method, called RSAE-VC (representation separation auto-encoder voice conversion) was proposed.The speaker’s personality characteristics in the speech were regarded as time invariant and the content information as time variant, and the instance normalization and activation guidance layer were used in the encoder to separate them from each other.Then the content information of the source speech and the personality characteristics of the target one was utilized to synthesize the converted speech by the decoder.The experimental results demonstrate that RSAE-VC has an average reduction of 3.11% and 2.41% in Mel cepstral distance and root mean square error of pitch frequency respectively, and has an increasement of 5.22% in MOS and 8.45% in ABX, compared with the AGAIN-VC (activation guidance and adaptive instance normalization voice conversion) method.In RSAE-VC, self-content loss is applied to make the converted speech reserve more content information, and self-speaker loss is used to separate the speaker personality characteristics from the speech better, which ensure the speaker personality characteristics be left in the content information as little as possible, and the conversion performance is improved.

A joint user association and dynamic resource allocation algorithm was proposed for the slicing scenario of ultra dense low earth orbit-radio access network (LEO-RAN) in order to address the efficient utilization of resources of the integrated terrestrial-satellite network for 6G .Considering the constraints of the minimum rate, maximum delay and resource proportion of different slices, a joint optimization problem of user association and resource allocation was established to maximize the weighted sum of the SE and the differentiated SLA of different slices as the optimization objective.A network slicing algorithm based on multi-agent deep deterministic policy gradient (MADDPG) was proposed to determine the proportion of slicing resources, a Lagrange dual based user association algorithm was proposed to determine the optimal user association policy and the resources were allocated to users by using the round-robin scheduling mechanism.The simulation results show that the proposed algorithm can effectively improve SE while satisfying the differentiated SLA of different slices.Compared with MADDPG-RA, MATD3-LG, MATD3-RA, MASAC-LG and MASAC-RA algorithms, the system utility of the proposed algorithm is improved by 2.0%, 2.3%, 5.7%, 8.7% and 9.4%, respectively.

Focused on the problem that the stream computing platform was suffering from the high energy consumption and low efficiency due to the lack of consideration for energy efficiency in designing process, an energy-efficient optimization strategy based on elastic data migration in big data streaming platform (EEDM-BDSP) was proposed.Firstly, models of the load prediction and the resource judgment were set up, and the load prediction algorithm was designed, which predicted the load tendency and determine node resource occupancy, so as to find nodes of resource overload and redundancy.Secondly, models of the resource constraint and the optimal data migration were set up, and the optimal data migration algorithm was proposed, which data migration for the purpose of improving node resource utilization.Finally, model of the energy consumption was set up to calculate the energy consumption saved by the cluster after data migration.The experimental results show that the EEDM-BDSP changes node resources in the cluster can responded on time, the resource utilization and the energy-efficient are improved.

Aiming at the problems of high-latency, high energy consumption, high bandwidth, and poor quality of experience (QoE) caused by emerging network media streaming business in mobile edge computing, a computing offloading algorithm based on QoE feedback configuration was proposed.Firstly, both preprocessing and priority were comprehensively considered to maximize network resource utilization.Meanwhile, different weights were assigned to the computation tasks for establishing a resource allocation relationship.Secondly, after comprehensively taking into account deadline, computing resource, power and bandwidth constraint, an QoE model was established where the optimization objective was the weighted sum of task delay, energy consumption and precision, and the method of Lagrange multipliers was utilized to solve the established model.Simulation results indicate that, compared with the deep reinforcement learning-based online offloading algorithm, the proposed algorithm can effectively optimize the resource allocation and better improve the QoE.

To address the problems of limited number of substitutable words and low watermark extraction efficiency in the existing natural language digital watermarking methods, a creative method based on context word prediction and window compression coding was proposed.Firstly, the contextual semantic features of each word in the original text were automatically learned through a neural network language model, and then the candidate word set for each word was predicted, thus the number of substitutable words that could be utilized for carrying watermark information was expanded.Meanwhile, considering the difference of the semantic impact caused by the substitutions of candidate words at different positions, the watermark information was embedded into each window containing several words, and the selection of candidate words for watermark embedding was optimized by the similarity between sentences before and after performing word substitutions.Finally, a semantic-independent window compression coding method was proposed, which encoded each window as appointed watermark information in terms of the character information of words contained in the window.So that during watermark extraction, the dependence on the original context at the position of word substitution was eliminated.The experimental results show that the proposed method greatly improves the watermark extraction efficiency with high embedding capacity and text quality.

In order to further improve the self-learning and noise reduction ability of speech enhancement model, a speech enhancement method based on multi-domain fusion and neural architecture search was proposed.The multi-spatial domain mapping and fusion mechanism of speech signals were designed to realize the mining of real complex number correlation.Based on the characteristics of convolution pooling of the model, a complex neural architecture search mechanism was proposed, and the speech enhancement model was constructed efficiently and automatically through the designed search space, search strategy and evaluation strategy.In the comparison and generalization experiment between the optimal speech enhancement model and the baseline model, the two indexes of PESQ and STOI increase by 5.6% compared with the optimal baseline model, and the number of model parameters is the lowest.

The unpredictable environmental changes cause large fluctuations in Wi-Fi signals, and it is difficult to quantify the static and behaviorally-aware dynamic components in the subcarriers, so it is not possible to accurately portray the fluctuating form of dynamic breathing characteristics.Based on this, a subcarrier mutual information breathing perception (SMIBP) system was proposed.Firstly, a form of inscribing dynamic component information (DCI) was proposed, and mutual information theory was utilized to extract the dynamic component information representing respiration in subcarriers.Then, analytic hierarchy process (AHP) was used to combine the subcarriers to maximize the dynamic characteristics of the respiratory signals to obtain the reconstructed DCI sensory base signals.Finally, the respiration rate was obtained by combining the wavelet transform and peak detection method, which revealed of a theoretical pathway representing the dynamic component of human respiration.Simulation results show that the proposed system can better characterize the dynamic breathing component in each subcarrier, and can significantly improve the sensing accuracy and range of Wi-Fi for small-scale actions in different scenarios.