FuseNet：应用于移动端的轻量型图像识别网络

为解决目前Transformer模型因其巨大的参数量和计算复杂度而无法应用在计算资源相对有限的移动设备中的问题,提出了一种适用于移动端的友好型轻量图像识别网络称为FuseNet。FuseNet利用卷积神经网络提取局部特征信息和自注意力机制擅长对全局信息进行建模的特性,将局部表示与全局表示两者的特性整合至一个特征提取模块,高效融合了两种不同结构的优点达到以较小的模型规模实现较高准确率的目的。实验证明不同参数规模的FuseNet可以在不进行预训练的情况下实现良好的性能,可以很好地应用在移动设备中。FuseNet-B在ImageNet-1K数据集上以14.8M的参数量实现了80.5%的准确率,性能表现超过了同体量的Transformer模型和卷积神经网络。     

Context-aware mobile service for routing the fastest subway path

The rapid growth of the IT industry during the last few decades has increased demands on mobile devices such as PDAs, cellular phones, and GPS navigation systems. With emerging concepts of context-aware computing, the mobile devices can provide mobile users with timely information by using not only common knowledge but also environmental context such as current time and location. Lately, the context-aware applications have been actively investigated and have been contributed to numerous application areas such as real-time electronic catalogues and navigation systems for tourists. In this paper, we propose a new context-aware application for finding the fastest subway route. We have developed the proposed application as an implemented system named Optimize Your Time System (OYT System, for short). A terminal device of the OYT System is equipped with a GPS receiver and the system’s server contains a timetable of all trains in a target subway system. On perceiving users’ context such as current time and location automatically from GPS, the OYT System can display the optimal route which takes the shortest time for the user to reach the specified destination. In this paper, we present details of the OYT System and some experimental examples.     

Intrusion detection for mobile devices using the knowledge-based, temporal abstraction method

In this paper, a new approach for detecting previously unencountered malware targeting mobile device is proposed. In the proposed approach, time-stamped security data is continuously monitored within the target mobile device (i.e., smartphones, PDAs) and then processed by the knowledge-based temporal abstraction (KBTA) methodology. Using KBTA, continuously measured data (e.g., the number of sent SMSs) and events (e.g., software installation) are integrated with a mobile device security domain knowledge-base (i.e., an ontology for abstracting meaningful patterns from raw, time-oriented security data), to create higher level, time-oriented concepts and patterns, also known as temporal abstractions. Automatically-generated temporal abstractions are then monitored to detect suspicious temporal patterns and to issue an alert. These patterns are compatible with a set of predefined classes of malware as defined by a security expert (or the owner) employing a set of time and value constraints. The goal is to identify malicious behavior that other defensive technologies (e.g., antivirus or firewall) failed to detect. Since the abstraction derivation process is complex, the KBTA method was adapted for mobile devices that are limited in resources (i.e., CPU, memory, battery). To evaluate the proposed modified KBTA method a lightweight host-based intrusion detection system (HIDS), combined with central management capabilities for Android-based mobile phones, was developed. Evaluation results demonstrated the effectiveness of the new approach in detecting malicious applications on mobile devices (detection rate above 94% in most scenarios) and the feasibility of running such a system on mobile devices (CPU consumption was 3% on average).     

基于改进YOLOv5的轻量化航空目标检测方法

为解决硬件平台资源受限条件下的实时航空目标检测需求,在基于改进YOLOv5的基础上,提出了一种针对移动端设备/边缘计算的轻量化航空目标检测方法。首先以MobileNetv3为基础搭建特征提取网络,设计通道注意力增强结构MNtECA (MobileNetv3 with Efficient Channel Attention)提高特征提取能力;其次在深度可分离卷积层增加1×1的卷积,在减少卷积结构参数的同时提高网络的拟合能力;最后对检测网络进行迭代通道剪枝实现模型压缩和加速。实验选取DIOR (Object Detection in Optical Remote Sensing Images)数据集进行训练和测试,并在嵌入式平台(NVIDIA Jetson Xavier NX)对轻量级模型进行推理验证。结果表明,所提出的轻量级模型大幅降低了参数和计算量,同时具有较高精度,实现了移动端设备/边缘计算的实时航空目标检测。     

Constrained temporal structure for text-dependent speaker verification

In the context of mobile devices, speaker recognition engines may suffer from ergonomic constraints and limited amount of computing resources. Even if they prove their efficiency in classical contexts, GMM/UBM systems show their limitations when restricting the quantity of speech data. In contrast, the proposed GMM/UBM extension addresses situations characterised by limited enrolment data and only the computing power typically found on modern mobile devices. A key contribution comes from the harnessing of the temporal structure of speech using client-customised pass-phrases and new Markov model structures. Additional temporal information is then used to enhance discrimination with Viterbi decoding, increasing the gap between client and imposter scores. Experiments on the MyIdea database are presented with a standard GMM/UBM configuration acting as a benchmark. When imposters do not know the client pass-phrase, a relative gain of up to 65% in terms of EER is achieved over the GMM/UBM baseline configuration. The results clearly highlight the potential of this new approach, with a good balance between complexity and recognition accuracy.     

Research on the strategy of mobile short video in product sales based on 5G network and embedded system

Mobile short video-based product sales sharing sites like YouTube and Tudor have many established user content for creating and distributing shares. The increasing number of mobile devices for product sales leads to the upcoming new 5G technology roadmap for embedded systems and 5G network connectivity. As these are the main sources of 5G information and online activities for consumers, mobile phone short films are rapidly being replaced by embedded systems. As the demand for more embedded system devices and applications continues to grow, supported bandwidth is also essential to meet this growing connection demand. The existing system does not allocate the product sales data upload bandwidth size. The system proposed here focuses on user upload bandwidth allocation, one of the basic resources of a short video sharing system with product details. Its allocation upload bandwidth Recurrent Neural Network (RNN) algorithm is proposed in a centralized or decentralized way and evaluated for balancing widely used strategies (equal allocation) and a mobile short video. Embedded systems are responsible for running professional product sales and control applications consistently and predictably. Development while using the microprocessor is also important. It increases the need to process product sales to handle the bandwidth, latency requirements, product sales data and data generated from multiple connected devices. It's a big challenge for the industry to data and data capabilities.     

Mobile Surface Reflectometry

We present two novel mobile reflectometry approaches for acquiring detailed spatially varying isotropic surface reflectance and mesostructure of a planar material sample using commodity mobile devices. The first approach relies on the integrated camera and flash pair present on typical mobile devices to support free‐form handheld acquisition of spatially varying rough specular material samples. The second approach, suited for highly specular samples, uses the LCD panel to illuminate the sample with polarized second‐order gradient illumination. To address the limited overlap of the front facing camera's view and the LCD illumination (and thus limited sample size), we propose a novel appearance transfer method that combines controlled reflectance measurement of a small exemplar section with uncontrolled reflectance measurements of the full sample under natural lighting. Finally, we introduce a novel surface detail enhancement method that adds fine scale surface mesostructure from close‐up observations under uncontrolled natural lighting. We demonstrate the accuracy and versatility of the proposed mobile reflectometry methods on a wide variety of spatially varying materials.     

A cache invalidation strategy for mobile networks

The diversity of services delivered over wireless channels has increased people's desire in ubiquitously accessing these services from their mobile devices. However, a ubiquitous mobile computing environment faces several challenges such as scarce bandwidth, limited energy resources, and frequent disconnection of the server and mobile devices. Caching frequently accessed data is an effective technique to improve the network performance because it reduces the network congestion, the query delay, and the power consumption. When caching is used, maintaining cache consistency becomes a major challenge since data items that are updated on the server should be also updated in the cache of the mobile devices. In this paper we propose a new cache invalidation scheme called Selective Adaptive Sorted (SAS) cache invalidation strategy that overcomes the false invalidation problem that exists in most of the invalidation strategies found in the literature. The performance of the proposed strategy is evaluated and compared with the selective cache invalidation strategy and the updated invalidation report startegy found in the literature. Results showed that a significant cost reduction can be obtained with the proposed strategy when measuring performance metrics such as delay, bandwidth, and energy.     

Multiobjective computation offloading for workflow management in cloudlet‐based mobile cloud using NSGA‐II

Cloudlet is a novel computing paradigm, introduced to the mobile cloud service framework, which moves the computing resources closer to the mobile users, aiming to alleviate the communication delay between the mobile devices and the cloud platform and optimize the energy consumption for mobile devices. Currently, the mobile applications, modeled by the workflows, tend to be complicated and computation‐intensive. Such workflows are required to be offloaded to the cloudlet or the remote cloud platform for execution. However, it is still a key challenge to determine the offloading resolvent for the deadline‐constrained workflows in the cloudlet‐based mobile cloud, since a cloudlet often has limited resources. In this paper, a multiobjective computation offloading method, named MCO, is proposed to address the above challenge. Technically, an energy consumption model for the mobile devices is established in the cloudlet‐based mobile cloud. Then, a corresponding computation offloading method, by improving Nondominated Sorting Genetic Algorithm II, is designed to achieve the goal of energy saving for all the mobile device while satisfying the deadline constraints of the workflows. Finally, extensive experimental evaluations are conducted to demonstrate the efficiency and effectiveness of our proposed method.     

Protecting lightweight block cipher implementation in mobile big data computing

The Mobile Big Data Computing is a new evolution of computing technology in data communication and processing. The data generated from mobile devices can be used for optimization and personalization of mobile services and other profitable businesses. Mobile devices are usually with limited computing resources, thus the security measures are constrained. To solve this problem, lightweight block ciphers are usually adopted. However, due to the easily exposed environment, lightweight block ciphers are apt to suffer from differential power attack. To counteract this attack, Nikova et al. proposed a provably secure method, namely sharing, to protect the cipher’s implementation. But the complexity of sharing method is so high, making this method not practical. To address this issue, in this paper, we propose a GPU-based approach of sharing a 4-bit S-box by automatic search. GPU is a promising acceleration hardware with powerful parallel computing. By analyzing the sharing method carefully, we devise an optimal approach, namely OptImp, that improves the performance massively. The experiment results show that the proposed approach can achieve up to 300 times faster than the original method. With our approach, the sharing method can be used to protect lightweight block ciphers in practice.     

Skeleton-based action recognition with temporal action graph and temporal adaptive graph convolution structure

Skeleton-based action recognition has recently achieved much attention since they can robustly convey the action information. Recently, many studies have shown that graph convolutional networks (GCNs), which generalize CNNs to more generic non-Euclidean structures, are more exactly extracts spatial feature. Nevertheless, how to effectively extract global temporal features is still a challenge. In this work, firstly, a unique feature named temporal action graph is designed. It first attempts to express timing relationship with the form of graph. Secondly, temporal adaptive graph convolution structure (T-AGCN) are proposed. Through generating global adjacency matrix for temporal action graph, it can flexibly extract global temporal features in temporal dynamics. Thirdly, we further propose a novel model named spatial-temporal adaptive graph convolutional network (ST-AGCN) for skeletons-based action recognition to extract spatial-temporal feature and improve action recognition accuracy. ST-AGCN combines T-AGCN with spatial graph convolution to make up for the shortage of T-AGCN for spatial structure. Besides, ST-AGCN uses dual features to form a two-stream network which is able to further improve action recognition accuracy for hard-to-recognition sample. Finally, comparsive experiments on the two skeleton-based action recognition datasets, NTU-RGBD and SBU, demonstrate that T-AGCN and temporal action graph can effective explore global temporal information and ST-AGCN achieves certain improvement of recognition accuracy on both datasets.

     

Improving mobile search using content enrichment

Providing an effective mobile search service is a difficult task given the unique characteristics of the mobile space. Small-screen devices with limited input and interaction capabilities do not make ideal search devices. In addition, mobile content, by its concise nature, offers limited indexing opportunities, which makes it difficult to build high-quality mobile search engines and indexes. In this paper we consider the issue of limited page content by evaluating a heuristic content enrichment framework that uses standard Web resources as a source of additional indexing knowledge. We present an evaluation using a mobile news service that demonstrates significant improvements in search performance compared to a benchmark mobile search engine.     

Context provenance to enhance the dependability of ambient intelligence systems

Ambient intelligence systems would benefit from the possibility of assessing quality and reliability of context information based on its derivation history, named provenance. While various provenance frameworks have been proposed in data management, context data have some peculiar features that claim for a specific support. However, no provenance model specifically targeted to context data has been proposed till the time of writing. In this paper, we report an initial investigation of this challenging research issue by proposing a provenance model for data acquired and processed in ambient intelligence systems. Our model supports representation of complex derivation processes, integrity verification, and a shared ontology to facilitate interoperability. The model also deals with uncertainty and takes into account temporal aspects related to the quality of data. We experimentally show the impact of the provenance model in terms of increased dependability of a sensor-based smart-home infrastructure. We also conducted experiments to evaluate the communication and computational overhead introduced to support our provenance model, using sensors and mobile devices currently available on the market.     

基于STM32的双向交互式车载通信系统设计

提出了一种基于STM32的车载通信系统。该系统由STM32 F103 ZET6硬件平台和μC/OS III嵌入式实时操作系统软件平台构成,采用 GPS定位技术获取车辆实时位置,CAN总线技术获取 ADAS设备的行车预警信息和车辆状态信息,应用3G无线网络传输技术上传行车预警特征信息,形成车辆与服务平台双向交互式通信。实验结果表明,该车载通信系统实现了车辆与网络服务平台间高速、可靠、实时通信,具有较高的实用价值。     

A robust web personal name information extraction system

Personal information extraction, which extracts the persons in question and their related information (such as biographical information and occupation) from web, is an important component to construct social network (a kind of semantic web). For this practical task, two important issues are to be discussed: personal named entity ambiguity and the extraction of personal information for a specific person. For personal named entity ambiguity, which is a common phenomenon in the fast growing web resource, we propose a robust system which extracts lightweight features with a totally unsupervised approach from broad resources. The experiments show that these lightweight features not only improve the performances, but also increase the robustness of a disambiguation system. To extract the information of the focus person, an integrated system is introduced, which is able to effectively re-use and combine current well-developed tools for web data, and at the same time, to identify the expression properties of web data. We show that our flexible extraction system achieves state-of-the-art performances, especially the high precision, which is very important for real applications.     

“Cache and Carry” for Location Management in Mobile Information Systems

Recent development of hardware technologies such as communication medium which advances from wired to wireless has led to the emergence of mobile information systems. A major problem in such a mobile information system is how to locate mobile clients. This is named the location management issue. Two major costs are involved in managing a mobile client's location: the movement cost and the locating cost. Past methods can only minimize one of the two costs, but not both. The major contribution of this paper is to present methods that minimize both costs simultaneously. Our performance analysis proves that the proposed methods are superior to the past ones.     

Unsupervised named-entity extraction from the Web: An experimental study

The KnowItAll system aims to automate the tedious process of extracting large collections of facts (e.g., names of scientists or politicians) from the Web in an unsupervised, domain-independent, and scalable manner. The paper presents an overview of KnowItAll's novel architecture and design principles, emphasizing its distinctive ability to extract information without any hand-labeled training examples. In its first major run, KnowItAll extracted over 50,000 class instances, but suggested a challenge: How can we improve KnowItAll's recall and extraction rate without sacrificing precision?This paper presents three distinct ways to address this challenge and evaluates their performance. Pattern Learning learns domain-specific extraction rules, which enable additional extractions. Subclass Extraction automatically identifies sub-classes in order to boost recall (e.g., “chemist” and “biologist” are identified as sub-classes of “scientist”). List Extraction locates lists of class instances, learns a “wrapper” for each list, and extracts elements of each list. Since each method bootstraps from KnowItAll's domain-independent methods, the methods also obviate hand-labeled training examples. The paper reports on experiments, focused on building lists of named entities, that measure the relative efficacy of each method and demonstrate their synergy. In concert, our methods gave KnowItAll a 4-fold to 8-fold increase in recall at precision of 0.90, and discovered over 10,000 cities missing from the Tipster Gazetteer.     

基于残差时域注意力神经网络的交通模式识别算法

交通模式识别是用户行为识别中的一个重要分支,其目的是对用户所处的交通模式进行准确判断。针对现代智慧城市交通系统对在移动设备环境下精准感知用户交通模式的需求,提出了一种基于残差时域注意力神经网络的交通模式识别算法。首先,通过具有较强局部特征提取能力的残差网络提取传感器时序中的局部特征;然后,采用基于通道的注意力机制对不同传感器特征进行重校准,并针对不同传感器的数据异构性进行注意力重校准;最后,利用具有更广感受野的时域卷积网络（TCN）提取传感器时序中的全局特征。采用数据丰富度较高的宏达通讯（HTC）交通模式识别数据集来对已有的交通模式识别算法和所提出的残差时域注意力模型进行评估,实验结果表明,所提出的残差时域注意力模型在对现代移动嵌入式设备的计算开销友好的前提下具有高达96.07%的准确率,且对单一类别均具有高于90%的召回率与精确率,验证了该模型的准确性与鲁棒性。所提模型可以作为一种支持移动智能终端运算的交通模式识别应用于智能交通出行、智慧城市等领域。