Network traffic anomaly detection is limited by the lack of annotation information in the traffic. This paper presents an unsupervised anomaly detection method based on score iterations that overcomes this limitation. An autoencoder based anomaly score iteration process was designed to learn generic anomaly features to determine an initial anomaly score. A deep ordinal regression model based anomaly score iteration process was then designed to learn discriminative anomaly features to further improve the anomaly score accuracy. Deep models, multi-view features and ensemble learning are also used to improve the detection accuracy. Tests on several datasets show that this method has significant advantages over other methods in the absence of annotation information and can be effectively applied to network traffic anomaly detection.

Quick response (QR)-in-QR attacks are a type of QR code hijacking. The scanner needs to first identify the finder patterns to determine the location of the QR code and the QR code needs to be surrounded by a quiet zone to help determine the location. Existing techniques cannot be used for actual attack scenarios due to the complex visual characteristics of the finder patterns and quiet zones. This paper presents an invisible QR-in-QR hijacking attack based on finder pattern modification and hidden quiet zones. The finder patterns of the malicious QR code can be modified to hide the malicious QR code for a targeted attack on the specified software. The quiet zones can be hidden to hide the position of the malicious QR code. Tests show that the invisible QR-in-QR hijacking attack method can implement effective attacks while hiding visual characteristics and can selectively attack WeChat and Alipay.

Existing federated learning models have problems due to malicious central servers and malicious participants publishing false data that poisons the model. A decentralized federated learning model was developed to address these problems by moving the aggregation work from the central server to the participants' computers. Each participant uses the aggregation algorithm to write the trained model parameters into the transaction and generates blocks that are then published to the blockchain network. A Byzantine fault-tolerant consensus algorithm based on model accuracy is used to build a consensus group and the nodes are dynamically joined by establishing a node information table. The results show that under the same conditions, compared with the traditional Byzantine fault-tolerant consensus algorithm, the throughput of the high-performance Byzantine fault-tolerant consensus algorithm based on model accuracy is increased by 60%, and the average system delay is reduced from 6 s to 1 s.

Current network security situation assessment methods have limited feature extraction capabilities and can be more efficient. This paper presents a network security situation assessment method that uses a parallel feature extraction network (PFEN) and an improved bi-directional gate recurrent unit (BiGRU). A deep learning model is designed with a PFEN and a BiGRU based on an attention mechanism (ABiGRU). The PFEN module has parallel sparse auto-encoders which identify key data out of the network traffic and integrate this data with the original features. Then, the ABiGRU module weights the key features through the attention mechanism to improve the model accuracy. The trained PFEN-ABiGRU is then applied to network threat detection. The model detection results are combined with a network security quantification method to calculate a network security situation index. Tests indicate that the PFEN-ABiGRU assessments have better accuracy and recall rates than other model assessment results.

Key node recognition methods for complex networks often have insufficient resolution and accuracy. This study developed a K-shell based key node recognition method for complex networks that first stratifies the network to obtain the K-shell (K_s) values for each node that indicate the influence of the global structure of the complex network. A comprehensive degree (CD) was then defined that balances the various influences of neighboring nodes and sub-neighboring nodes. A dynamic adjustable influence coefficient, μ_i, was also defined. Nodes with the same K_s but larger comprehensive degrees are more important. Tests show that this method more effectively identifies key nodes than several classical key node recognition methods and a risk assessment method, and has high accuracy and resolution in different complex networks. This method provides network node risk assessments that can be used to protect important nodes and to determine the risk disposal priority of the network nodes.

A multi-key privacy-preserving decision tree evaluation (MPDE) scheme was developed to protect the privacy of decision tree data and models in machine learning and to reduce the computational and communications overhead. A distributed two-trapdoor public-key crypto (DT-PKC) was used to encrypt all the data. A secure addition- across-domains protocol was then used to add two ciphertexts from different public key cryptography systems. In addition, the original security comparison protocol was improved to support multi-user, multi-key systems to protect the privacy of the requested information, classification results and decision tree model. A trusted third party key generation center was introduced to reduce the communication overhead between entities which is completely offline after the key distribution. A service agent was then used to interact with the cloud server instead of the users which reduced the communications overhead between the user and the cloud server. Security and performance analyses show that the scheme is efficient and ensures privacy. Simulations show that the scheme has less computational overhead than previous schemes.

A ground source heat pump system was designed for heating embankments to eliminate frost heave. Tests showed that the device can automatically output heating temperatures of 40, 50, or 60℃ with coefficients of performance greater than 3.0. The heating influenced a 0.76 m radius region on the 1^st day, 1.64 m on the 5^th day, and 2.30 m on the 10^th day. A case study showed that the central freezing depth of an embankment was 0.89 m for standard conditions but was reduced to less than 0.2 m with artificial heating. The soil temperature rise and the heat diffusion both increased with increased heating temperature. Rapid thawing and emergency rescues during frost heave conditions require that the heat pumps have longitudinal spacings along the embankment of 2.0-4.0 m and heating capacities of 1.0-2.0 kW.

The use of a phase change material (PCM) encapsulated in a steel ball in place of the coarse aggregate in concrete can improve the energy density and heat transfer in the energy pile which will reduce the underground space needed for the heat transfer. Cooling-heating loads are used in a traditional concrete energy pile and a PCM energy pile in a container containing unsaturated clay to experimentally study the thermal response of the piles and surrounding soil. The results show that the temperature influence range in the soil surrounding the phase change pile extends out to about 1.5 times the pile diameter during the cooling-heating processes with a larger temperature difference between the PCM energy pile inlet and outlet than with the traditional concrete energy pile, which indicates a larger heat transfer rate. The temperature differences in both the PCM pile and the traditional pile during heating are less than during cooling which shows that the heat transfer rates during cooling are larger than during heating for the same flow conditions. The results also show that the PCM increases the uneven temperature distribution during the cooling in the vertical and horizontal directions in the pile. In addition, irreversible settling of unsaturated clay is observed at the soil surface due to temperature induced soil consolidation and drainage.

The effects of groundwater seepage on the heat transfer and thermo-mechanical coupling characteristics of an energy pile were analyzed using a coupled thermal-mechanical numerical model of the energy pile with seepage. The model shows how the groundwater seepage affects the thermodynamic properties of the energy pile during summer. The results show that in the summer mode the heat transfer of the energy pile with a horizontal seepage velocity of 60 m/a is 1.34 times larger than without seepage. The temperature rise in the pile is then reduced by 9.12%. The groundwater seepage reduces the variations of the pile body displacement, axial force, and the pile side friction while the energy pile rapidly becomes stable. In addition, the groundwater seepage reduces the effect of soil heat in the upstream seepage, but increases the influence of the soil heat in the downstream seepage.

As a multimodal task, visual question answering (VQA) requires a comprehensive understanding of images and questions. However, conducting reasoning simply on images and questions may fail in some cases. Other information that can be used for the task, such as image captions and external knowledge base, exists. A novel approach is proposed in this paper to incorporate information on image captions and external knowledge into VQA models. The proposed approach adopts the co-attention mechanism and encodes image captions with the guidance from the question to utilize image captions. Moreover, the approach incorporates external knowledge by using knowledge graph embedding as the initialization of word embeddings. The above methods enrich the capability of feature representation and model reasoning. Experimental results on the OKVQA dataset show that the proposed method achieves an improvement of 1.71% and 1.88% over the baseline and best-reported previous systems, respectively, which proved the effectiveness of this method.

The important considerations for data annotation are the annotation data quality and the annotation cost. Data annotation in natural language processing usually first uses automated model annotation followed by human corrections to reduce the cost. There have been few studies comparing the effects of different annotation approaches on the annotation quality and cost. This study uses a mature annotation team completing a dependency tree annotation as a case study. This study compares three data annotation approaches using model annotation followed by human corrections, double-blind annotation, and human-model double-blind annotation that is the fusion of the first two approaches. The human-model double-blind annotation effectively combines the advantages of model annotation followed by human corrections and double-blind annotation to reduce the annotation cost and then to improve the annotation quality by eliminating the identification tendency problem.

In software-defined networking (SDN), the overlapping among the matching fields of rules complicates flow table updates. One update often triggers the movement of multiple ternary content addressable memory (TCAM) entries, which increases the update times. In addition, the TCAMs in existing SDN switches are mostly designed with a single port. The TCAM update suspends the packet lookup which affects the packet forwarding in the data plane. Therefore, how to achieve fast update while supporting wire-speed packet lookup is an important research topic to improve network performance. This paper presents a TCAM-based SDN switch with a flow table update system. When multiple network application updates are integrated at the front end and simultaneously sent to the switch, the system can efficiently detect the dependencies between rules and prioritize the rules which need to be quickly updated, so that they can respond quickly. The update algorithm does not need to block the TCAM search operation and can provide interleaved execution of the packet lookups and rule updates. Tests show that these scheduling strategies improve the system performance by balancing the lookup first algorithm and the update first algorithm.

Multiprotocol packet forwarding in software-defined satellite networks involves large flow tables and expensive storage in onboard devices. A multiprotocol flow table architecture was developed to reduce information storage using a 2-dimensional expanded-field search (2D-EFS) algorithm for the limited resources of satellite networks. The 2D-EFS algorithm generates multiple flow tables by progressively merge fields to compress storage for flow table initialization and flow entry updates. Simulations show that the storage compression efficiency for flow table initialization can reach 86%, which is close to the global optimal and which outperforms existing single-protocol algorithms. The algorithm achieves an average storage compression of 76% for flow entry updates and has the shortest run time and better overall performance than existing single- protocol algorithms.

Ultra wide band (UWB) positioning technology establishes a mathematical model of the target location based on distance parameters. The problem of accurately solving the tag point positioning is transformed into an optimization problem of nonlinear equations due to ranging errors. Existing optimization methods may fail when certain distance errors are large. Therefore, this paper proposes an optimization method based on redundant distance screening by considering the distances between multiple positioning tags as constraints, filtering the redundant distances by weighting, and using a gradient descent method to optimize the initial value calculated by the Caffery method. In simulation experiments, the positioning error of this method is only 70% of the Caffery-Taylor (CT) method. In real data experiments, this method outperforms the CT method in terms of the optimization effect.

Supercomputers provide enormous computing power for large applications. Traditional supercomputers have mainly targeted scientific computing problems. However, other applications have new requirements for the both supercomputer software and hardware designs. The New Generation Sunway supercomputer has an inefficient memory allocator when running in the dynamic mode. This study develops an efficient memory allocator, SWAlloc, that reduces the memory allocation time of the brain scale pretrained model training framework, BaGuaLu, by up to 75 839 times. Evaluations using PARSEC also show that SWAlloc can speed up the memory allocation by up to 51 times (36% on average). SWAlloc has been deployed on the New Generation Sunway supercomputer for use by various large applications, including SWPytorch and SWTensorFlow.

Human-machine conversation plays an important role in natural language processing and artificial intelligence. Human-machine conversation can be divided into the question answering system, task-oriented conversation, and chatting system according to the purpose of use. Among them, the chatting system usually requires higher personification. Based on the sequence transformation model of the long short-term memory network, the topic network is introduced in this study to explicitly extract the scene and topic information from the conversation, and this higher-level feature, which does not change with the word order, is inputted to the structure of the conversation model to guide the decoding and prediction processes together with the attention mechanism. Because of the difficulty of obtaining the topic information in advance, the topic network is modeled as an unsupervised learning structure. Thus, the training process needs to be divided into three steps. The experimental results show that the model can significantly improve the quality of the chatting system with appropriate training methods and structural parameters.

In recent years, the rapid development of generative adversarial networks (GAN) has made synthesized images more and more realistic, which poses great threats to individuals and society. Existing research has focused on passively identifying deepfakes, but real-world applications are usually insufficiently general and robust. This paper presents a method for deepfake provenance and forensics. Deepfakes hide secret information in facial images to track the source of the forged image. An end-to-end deep neural network was designed to include an embedding network, a GAN simulator, and a recovery network. The embedding network embeds the secret information in the picture while the recovery network extracts the information. The GAN simulator simulates various GAN-based image transformations. The average normalized cross correlation coefficient (NCC) of the restored images after tampering with known GANs is higher than 0.9 and the average NCC reaches around 0.8 with tampering by unknown GANs, which shows good robustness and generalization. In addition, the secret embedded information is well concealed and the average peak signal to noise ratio (PSNR) is about 30 dB.

Computer numerical control (CNC) machine tools are typical repairable electro-hydraulic mechanical systems. Repair activities do not return the machine tool to like-new conditions, but to some intermediate state of general repair. A superposed log-linear proportional intensity model (S-LPIM) was used here to model the general repair state of machine tools using a least squares parameter estimation method. A failure intensity model was used to calculate a reliability evaluation index for the CNC machine tool with the failure probability and repair efficiency of the CNC machine tool evaluated based on the likelihood ratio test method. Tests of the reliability of three CNC machining centers verified that the S-LPIM model can fit the failure rate bathtub curve of machine tools and quantitatively describe the effect of maintenance activities on the machine tool reliability.

A stiff yet flexible paint spraying robot with a 3-DOF parallel mechanism and a rotating joint was mounted on a mobile platform to facilitate automobile painting in a repair shop. The kinematic method was used to develop a dynamic model of the 3-DOF parallel mechanism using the virtual work principle. The inertia matrix including the gravitational term was then used to evaluate the dynamic fluctuations with a global index defined to quantify the performance fluctuations. The effectiveness of the evaluation index was verified by the differences in the robot driving forces for various geometric and inertia parameters. The dynamic performance fluctuation index reflects the spatial fluctuations of the robot in the workspace and can be used to optimize the robot design and control.

Industrial software is a key force supporting the development of domestic small and medium-sized enterprises. Industrial software packages contain a large amount of knowledge related to manufacturing processes, but little of the knowledge embedded in these software packages has been extracted and put into a knowledge base. This paper presents a knowledge extraction model that combines neural networks and natural language processing. The model includes text representation, entity recognition, and relationship extraction. The extracted entities and relationships are modeled on a knowledge graph, while related concepts in the software are defined through ontology modeling. The ontology model concepts are mapped to graph data to improve data retrieval and modeling capabilities and the data can be stored in the knowledge base with long term. The results show that this method can build an industrial software knowledge base which will play an important role in integrating manufacturing knowledge.