Large-scale Deterministic Network (LDN) is presented to achieve deterministic transmission in the Wide Area Networks (WAN), which provides bounded delay and jitter with the Cycle Specified Queuing and Forwarding (CSQF) and shaping mechanisms. However, routing of time-sensitive flows in LDN still proves to be challenging. As flows’ rates are adjusted by the shapers at ingress nodes, leading to uncertainty of delay and bandwidth constraints, traditional routing algorithms used to solve the Multi-Constrained Path (MCP) problem cannot be directly applied in LDN, which might encounter the early rejection of flows and have overlong runtime for online scenarios. In this paper, we propose a Heuristic Online Algorithm based on a PathBook called PB-HOA, which computes paths in advance and uses a metric at the path level that considers the variable flows’ rates to select routes. Our experiments show that PB-HOA can successfully generate routes for most flows at a fast execution speed, confirming its feasibility for routing in LDN.
Successful exchange of cooperative awareness messages(CAMs) among vehicles is essential in improving transportation safety for intelligent transportation systems (ITS). This paper presents performance enhancements of C-V2X CAM delivery with i) overlapped virtual cells with orthogonal sub-channel allocations, ii) extended MIB (master information block) with resource usage bitmap, and iii) smart roaming across overlapped virtual cells. In the proposed scheme, virtual cells are configured with 3 ~ 4 orthogonal sub-channels for up to 400 vehicles. The virtual cell manager (VCM) continuously monitors the usages of resources (sub-channels and sub-frames) for CAM broadcasting by vehicles, configures an extended MIB with resource usage bitmap, and periodically broadcasts the extended MIB with a bitmap. By checking the extended MIB with the resource usage bitmap, each vehicle can easily avoid collisions by selecting unused resource. The smart roaming among overlapped virtual cells guarantees seamless CAM deliveries even in the boundaries of virtual cells. The performance of the proposed scheme has been analyzed with NS-3 simulation, and measured results of CAM delivery rates are more than 99 percent in highly congested metropolitan areas.
Optical networks play a critical role in modern information society. Statistics have revealed that around 50% of optical service faults stem from "errors" in the optical transmission system links, which are monitored through frame overheads. Unfortunately, current maintenance methods are insufficient in accurately predicting these errors, leading to potential disruptions in network services. To address this issue, this study introduces the Gini importance-based feature selection approach and maximum likelihood-based logistic regression algorithm for predicting optical transmission system link performance degradation. To facilitate the application of this model in real-world networks, the Network AI (NAI) three-level intelligent architecture is proposed, and an optical transmission performance degradation prediction system is developed based on this architecture. Experimental results from the network experiment demonstrate the effectiveness of the proposed performance degradation model and NAI-based system.
Deep learning (DL) introduces a novel data-driven programming paradigm, where the system logic is constructed through data training. However, this approach poses challenges in terms of system analysis and defect detection. To address this issue, recent efforts have focused on testing deep learning systems using intuitive criteria known as Neuron Coverage (NC) and its variants. These criteria measure the activation proportion of neurons in a neural network. Unfortunately, recent DL applications have largely overlooked the importance of context during testing. To address this gap, this paper first incorporates the context of the DL pipeline deployed before test execution, such as medical diagnosis and Android Malware detection. Next, we formulate structural coverage criteria to guide test suite generation based on the properties of DL pipeline in different contexts. Furthermore, we proposed a coverage-guided search algorithm to efficiently generate test suites. Experimental results highlight the effectiveness of our approach in uncovering numerous erroneous behaviors in contexts such as medical image diagnosis and Android malware detection. This approach significantly enhances the robustness of DL models by considering the structural aspects of the DL pipeline.
In this paper, we present an Improved Transformer model and a self-supervised framework for detecting anomalies in satellite telemetry data. Previous approaches based on RNN and LSTM variants have shown some progress in anomaly detection, but still require improvement in detection performance and context parallelism. To address these challenges, we propose an innovative Improved Transformer model that incorporates correlation modeling, long-range modeling, and parallel reasoning capabilities. This model features a token split structure, position embedding, and mask unit, which enhance its overall performance. Our self-supervised framework consists of two stages: pre-training and fine-tuning. We conduct extensive experiments on NASA’s public dataset and demonstrate the effectiveness of our proposed framework. Using our self-supervised approach, we achieved a significant 95.7% (3-class) F1-score on the validation set, outperforming supervised counterparts that utilized the same training data.
Bitcoin transactions are created through the concept called Unspent Transaction Output (UTXO). Users put their own UTXOs as inputs into a transaction for Bitcoin transfer and create multiple outputs, each specifying the recipient’s wallet address and the amount to be sent. UTXO refers to an output that has not been used as an input for any transaction yet and each UTXO can only be used as an input once. However, attempting to use a UTXO more than once is called a double-spending attack. Although double-spending in Bitcoin is ultimately impossible due to the system structure, it can occur when a transaction is deemed confirmed and off-chain goods or services are provided before sufficient transaction finality is guaranteed. We consider an attempt of double-spending attack when a UTXO used as an input in a transaction for payment exists together with another transaction on the Bitcoin network that uses the same UTXO as an input. In previous research, we randomly deployed observer nodes on the Bitcoin network and proposed a method to detect double-spending attacks using transaction data in the memory pool and a graph neural network model. In this paper, we analyze the impact of adding observer nodes to the Bitcoin network on the performance of graph neural network-based Bitcoin double-spending attack detection. We conducted experiments to examine the performance differences among three strategies for adding observer nodes. However, it was difficult to compare clear differences due to the performance degradation of the model caused by the differences in graph structure between datasets. Therefore, we provide an analysis of the causes and suggestions for improvement.
The traditional monolithic system architecture faces challenges such as low development efficiency, high maintenance costs, and limited scalability. Refactoring it into a microservice architecture improves efficiency and scalability. Database decomposition is crucial in this process to achieve independence and decoupling between microservices. A decomposition method for relational database is proposed in this paper, which combines static analysis and dynamic analysis to build the incidence matrix between tables, and then clusters the incidence matrix by spectral clustering algorithm to obtain database decomposition results, and automatically matches the decomposition results with microservices. Based on the above mentioned method, an automatic relational database decomposition tool has been designed and implemented. Experimental verification has shown that the tool can efficiently achieve relational database decomposition, thereby enabling support for microservices refactoring to legacy systems.
The demands of today’s 5G mobile network, especially low latency and high bandwidth, are a big challenge for the 5G Core (5GC) provider. The most critical user data packet handler in the 5GC Network Function (NF) is the User Plane Function (UPF), which is responsible for moving data from the user equipment to the destination data network, and vice versa. Existing work mainly focuses on implementing UPF using the key technologies of high-speed data processing. In this paper, with a mobile core provider called free5GC for a stand-alone (SA) 5G network, we share our experience with the implementation of UPF by using a programmable hardware appliance, which can offer more Tbps compared to the implementation of software UPF that can offer only a few hundred Gbps. For that, we demonstrate how to build up a more flexible architecture of UPF by using the Software-Defined Networking (SDN) concept due to the opacity of protocol specification. We split the UPF control signal implementation into a software application, and user data packet processing into a programmable hardware appliance. We also show how to integrate a number of current UPF data plane free5GC implementations such as Data Plane Development Kit (DPDK), Linux kernel module, and SmartNIC. Furthermore, we analyze and make use of microservices to support the specific features of the UPF data plane that cannot be implemented in a programmable hardware appliance. We tested our free5GC mobile network and the new UPF design architecture that can run on a real programmable hardware appliance from Accton CSP-7551. The evaluation results show that our programmable user plane can reach the line rate.
With the increase of 5G devices and the continuous expansion of servers in the Internet, network service providers lack low-cost technologies to accurately predict changes in network traffic. Accurate prediction of network traffic changes is of great significance to network operators’ resource management, traffic engineering, capacity planning, and service quality. However, the network state is not always in a state of steady change in the real environment, but in a state of fluctuation, so the deep learning model close to the real situation must be nonlinear. Network State Prediction with Attention-Based Graph Convolutional Network are used to learn complex topological structures to capture spatial dependencies, and show superiority in handling complex and non-linear graph-structured data. In this paper, we propose an attention- based graph convolutional neural network network state prediction technique. A multihead attention mechanism is introduced in the propagation layer, so that the central node features can be differentiated in the attention of neighboring nodes during the aggregation process. In extensive experiments on real network traffic, it is shown that the proposed network state prediction method outperforms previous methods .
There is a need for transport synchronization reference signal across an optical network for 5G RAN. The traditional optical transport network (OTN) in case of transparent transporting a timestamp signal is not guaranteed to keep its accuracy. This experiment connected a precision time protocol (PTP) grandmaster and a PTP time slave clock via optical transport network without any vendor-specific encapsulation or operation for time transfer. Timing signals using ITU-T G.8275.1 were sent in the two-way operation. We found that the asymmetry of the PTP time transfer over OTN resulted in variant time transfer errors, even there is no other non-synchronization-related data loading in the same optical channel. We have run simulations of various OTN configurations in the laboratory. Thus, we can conclude with some certainty that similar results will be applicable if this technique are used to transfer time and frequency with high accuracy.
The failure of edge devices in the IoT will affect the use of IoT applications. The introduction of the federated learning can train efficient models for devices under the premise of protecting privacy. However, current solutions rarely focus on the problem of data heterogeneity on IoT devices. In this paper, we introduce two personalized federated learning algorithms to implement intrusion detection models, which aim to solve data heterogeneity. We perform diverse partitions on the IoT dataset to simulate data heterogeneity on devices. Our experiments show that the proposed models have high performance in detecting attacks under various data distributions. Under the Non-IID setting, the test accuracies of our models are 95.5% and 93.4%, which are 8.4% and 6.3% higher than the model using traditional federated learning (FedAvg), respectively.
The emerging reconfigurable circuit technology, which can establish circuit connections among switches, has been proposed as a promising paradigm for datacenter networks. This paper investigates how to accelerate the non-preemptive multicast flows in a demand-aware manner in reconfigurable datacenter networks. Firstly, the problem of scheduling the circuit switch to minimize the average completion time is introduced and proved to be NP-hard. A two-round matching algorithm is proposed to handle the conflicts between different multicast flows under the bandwidth constraint. The approximation ratio of the proposed algorithm is proved to be $2\sqrt {2n} $, where n denotes the number of Top-of-Rack (ToR). Finally, we have demonstrated that the proposed algorithm can effectively reduce the average completion time of the flow compared to state-of-the-art algorithms through extensive simulations.
Millimeter wave (mmWave) and terahertz (THz) communication protocols employ extensive antenna arrays to ensure reliable signal reception. However, aligning these narrow beams incurs a significant cost in beam training, which increases proportionally with the number of antennas. While beam prediction methods based on semantics from the RGB images demonstrate initial feasibility, they are still low on accuracy. Motivated by the fact that the recently introduced Segment Anything Model (SAM) can generate very accurate masks, which can be considered semantic information in this domain, this study proposes a beam prediction solution based on the masks of SAM. The SAM can extract more accurate semantic data from visual sources. Then instead of using the RGB images directly, the semantics images have been used to predict the beamforming vector using a lightweight LeNet5 model. The experimental results show that the SAM-based proposed method can perform significantly better than the two state-of-the-art deep learning models. The proposed solution method can achieve near 100% in top-5 beam prediction accuracy in real-world communication scenarios.
The most significant issue with QR code is that the information encoded within them cannot be recognized visually. To address the issue, we propose a QR code that superimposes text, which we term the "Text QR code". The proposed QR code can be recognized by humans as text by inserting characters onto the black and white modules. In this paper, we present an algorithm for the QR code to ensure document readability.
To ensure efficient network operations, internet service providers need to take management actions such as deploying, scaling, and live migrating new virtual network functions (VNFs) in response to user demands and changes in network environments. However, relying on manual measures to handle such changes can result in longer response times and additional costs in the event of problems, compared to proactive management measures. Therefore, it is necessary to analyze network traffic and predict future traffic to proactively manage services and systems. This paper proposes a network traffic prediction model as a preliminary study for proactive service function chaining (SFC) configuration and VNF scaling research and performs auto-scaling of VNFs in a simulation environment. The proposed model uses a state-of-the-art algorithm Temporal Fusion Transformer (TFT), which considers the heterogeneity of data in time series data prediction, to predict actual traffic measurement in an OpenStack environment. Then, it shows the results of auto-scaling in the simulation environment based on the threshold value. The TFT model showed a maximum reduction of 94% in Mean Squared Error (MSE) compared to Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models in two real measurement datasets. Also, we were able to confirm that the use of traffic predicted by TFT can reduce the scaling frequency by up to 76%.
The current intelligent connected vehicles (ICV) often need to share the detected intrusion events to the cloud for further collaborative investigation. However, it was found that the high volume of false alarms will bring congestion to the upstream channel, which will exhaust the bandwidth and even deny other legitimate data sharing. This paper therefore is motivated to use the machine learning-based intrusion detection approach to increase the detection performance, in particular, to reduce the false positive rate. With the effective feature selection over the datasets, our approach yields a higher detection performance and lower computational cost. Further, the experimental results show that our approach has a lower false positive rate than that of the previous works on the common datasets.
Recently, as the fields that utilize personal information have diversified, digital Identification and security solutions are being provided along with the development of the digital ecosystem. However, the problem of personal information being exposed due to the use of malicious codes and programs with weak security is constantly occurring. In a situation where cases of abuse of sensitive privacy data such as medical data as well as financial damages by unauthorized theft of exposed personal information are spreading, owning and managing your own personal data to solve this important problem A technology that can solve this is attracting attention. In this paper, the history and disadvantages of the identity authentication system used by existing companies and institutions are explained, and the self-sovereign identity card proposed to overcome them is explained. Next, we introduce a decentralized model designed to protect personal information by storing identity information and medical data on individual devices and IPFS with a decentralized structure using blockchain.
Work automation using RPA tools is rapidly becoming more widespread. When it comes to implementing horizontal expansion or maintaining existing RPA scenarios, operators are first required to read and interpret the RPA scenarios. However, it is typically quite difficult to get a complete picture of the scenario simply by reading the scenario due to a lot of works and specialized skills required. In this paper, we propose a method to automatically generate explanatory sentences in natural language using logs generated on the basis of RPA scenarios so that operators can read them directly and understand the content. However, in this paper only Japanese is aimed at. Subjective evaluations conducted with several participants in our previous study [4] have shown that sentences generated by the proposed method can be utilized for the interpretation of RPA scenarios. However, this study did not specify how to obtain nouns (operation points), which are necessary elements for generating explanatory text. Therefore, we developed a new method to acquire nouns (operation points), generated explanatory sentences based on them, and conducted evaluation experiments. Our findings demonstrate that the proposed method can generate explanatory sentences from operation logs that can be utilized for interpreting RPA scenarios.
Federated learning (FL) advances the field of distributed machine learning for facilitating the privacy-preserving management of edge devices and central servers. However, the majority of data among edge devices is non-IID (not Independent and Identically Distributed), making it challenging to achieve edge intelligence. In this work, we attempt to mitigate the above issue through knowledge distillation (KD) by sharing knowledge between the central server and edge devices. Specifically, we first investigate where (i.e. which feature layer) to conduct the distillation for knowledge sharing. We find that setting the feature layer to that before the classification head yields superior performance. Moreover, we investigate how to conduct the KD in terms of loss choices. We test various types of losses for enhancing the knowledge sharing and find that Center Kernel Alignment (CKA) achieves the best performance among the investigated loss metrics. Overall, this work sheds new light on where and how to perform KD in FL. Experimental results on MNIST and Fashion-MNIST demonstrate that our finding yields a performance gain of at least 4%.
The current Internet infrastructure and technology system faces major challenges in intelligent, diversified, personalized, highly robust, efficient and other aspects, and it is urgent to reform the network infrastructure. Polymorphic smart network (PINet), as an open network architecture based on full-dimensional definable technology, has been studied by more and more scholars. This paper first introduces the architecture of polymorphic smart network, then introduces the key technologies and related research of polymorphic smart network, then describes its simulation environment, and finally expounds the existing problems and possible breakthrough technologies and key points of PInet.
