Search

Aggressive data augmentation is a key component of the strong generalization capabilities of Vision Transformer (ViT). One such data augmentation technique is adversarial training (AT); however, many prior works have shown that this often results in poor clean accuracy. In this work, we present pyramid adversarial training (PyramidAT), a simple and effective technique to improve ViT's overall performance. We pair it with a "matched" Dropout and stochastic depth regularization, which adopts the same Dropout and stochastic depth configuration for the clean and adversarial samples. Similar to the improvements on CNNs by AdvProp (not directly applicable to ViT), our pyramid adversarial training breaks the trade-off between in-distribution accuracy and out-of-distribution robustness for ViT and related architectures. It leads to 1.82% absolute improvement on ImageNet clean accuracy for the ViT-B model when trained only on ImageNet-1K data, while simultaneously boosting performance on 7 ImageNet robustness metrics, by absolute numbers ranging from 1.76% to 15.68%. We set a new state-of-the-art for ImageNet-C (41.42 mCE), ImageNet-R (53.92%), and ImageNet-Sketch (41.04%) without extra data, using only the ViT-B/16 backbone and our pyramid adversarial training. Our code is publicly available at pyramidat.github.io., Accepted to CVPR22 (oral, best paper finalist). 33 pages, including references & supplementary material

In this article, an event-triggered state estimation problem for wireless sensor network systems affected by random packet losses and correlated noises is considered. A set of independent Bernoulli variables are used to describe the random packet losses in the measurement transmission. An event-triggered transmission strategy is introduced to decrease limited network bandwidth consumption, and the measurement noise is correlated with the process noises of the same moment and the previous moment. Event-triggered estimator of process noises under the linear minimum variance criterion is derived. Then, an event-triggered state estimation algorithm related to the packet loss rate, noise correlation coefficient and triggering threshold is designed. Sufficient conditions are provided to guarantee convergence of the estimation error covariances of the proposed estimator. Finally, comparative simulation verifies the effectiveness of our algorithm.

The static and dynamic characteristics of the 15 kV 4H-silicon carbide (SiC) p-channel insulated gate bipolar transistor (IGBT) are investigated by the numerical simulation. The effect of the JFET width $(L_{\mathrm{JFET}})$ on the blocking voltage, the on-state performance and the capacitance characteristic are analyzed particularly. As the $L_{\mathrm{JFET}}$ is 10 µm, a high blocking voltage of -15.94 kV and a forward voltage drop of -7.5 V at the forward current of -10 A can be achieved. Meanwhile, the $C_{\mathrm{ge}}$ is 1.4 nF and the $C_{\mathrm{gc}}$ is about 4 pF at the $V_{\mathrm{ce}}$ of -5 kV. Furthermore, the turn-on time is about 120 ns and the turn-off time is about 620 ns at -5 kV voltage, respectively.

In this paper, the clamped inductive turn-off ruggedness of a novel Trench Insulated Gate Bipolar Transistor (TIGBT) at overcurrent condition is studied by numerical simulation. This proposed structure is optimized by using split gate structure to improve the turn-off reliability. Simulation result shows that the maximum turn-off critical current of the proposed TIGBT increases for about 42.8% compared with the conventional TIGBT structure. The main reason is that the electric field of the proposed structure is redistributed by RESURF effect at the trench bottomed area, resulting to attenuate the current filament at local area. In addition, the hole which concentrated at the trench bottom area is less because the bottom part of split gate polysilicon is shorted with the emitter. Therefore, the local dynamic avalanche effect which triggered by the excessive carriers concentration and high peak electric filed accumulation is weakened. The turn-off ruggedness of the proposed structure is enhanced effectively and the electrical parameters are not compromised but even better compared with conventional structure.

With the continuous development and advancement of Internet technology, building an information management mechanism and a smart campus control model is an inevitable trend in keeping with the times. Among them, the decision support system can provide a more convenient management path for construction work and take advantage of Internet of Things technology and RFID radio frequency technology to ensure that big data management is more convenient and effective. This paper analyzes the characteristics of the smart campus and the overall framework of the decision support system, while focusing on the design and implementation process of the smart campus decision support system for reference only.

The current campus network has become the basic equipment in campus construction. The safe and reliable campus network not only can utilize various resources inside and outside the schoolreasonably, safely and effectively, but also improve the working environment of the campus network, facilitate the management and maintenance of the campus network, and improve the reliability of the campus network. Therefore, it is especially important to build a reasonable campus network security system; Based on the security design in the network construction of a university campus network, this paper analyzes comprehensively the possible security problems, the current security status of campus network and the needs of campus network security. Based on the existing information organization structure of colleges and universities, this paper uses various management methods and technical methods to divide the campus network into various functional area modules, and deploys corresponding solutions in each regional module to solve the network security existing in the campus network. The problem is ultimately to protect the online system resources and services, and to alleviate the security threats of the campus network.

In this paper, the facial expression recognition(FER) application is achieved by integrating Machine Learning and High Performance Computing (HPC). The convolutional neural network(CNN) will be trained utilizing the XSEDE high-performance computing resources. To observe the performance of HPC, CNN model has been trained with different values of parameters and their training times are provided. Compared to an ordinary computer, HPC with CPU and Google Colaboratory with GPU, this research explores the highest work efficiency of HPC with GPU with respect to operating time.

Dangerous situations such as children are left in vehicles, are dropped off at wrong stops, or take on wrong school buses usually caused by the negligence of drivers. This paper presents a real-time intelligent in-vehicle monitoring system that can count and recognize people as well as alert drivers if such improprieties or potential dangers happen. The system uses HOG-based face detector from Dlib library to obtain face counting function. Face recognition is achieved through two steps, facial feature extraction and face identification. The ResNet is used in facial feature extraction. It transforms an aligned face into a 256-dimensional vector, a Euclidean facial embedding. In face identification, labeled faces will be transformed to facial embeddings first. Then k-nearest neighbor classifier (kNN) is adopted to identify people using such facial embeddings. The simulation on ChokePoint dataset is tested and the average accuracy is 93 percent.

In this paper, a novel CSTBT with very narrow mesa structure is proposed by numerical simulation. The floating pwell region in the proposed structure is connected with the cathode through the very narrow mesa structure to deplete the very high dose concentration of the CS layer. Therefore, the conductivity modulation effect of the proposed structure can be enhanced significantly. It can help the proposed structure to achieve ultra-low on-state saturation voltage closed to the theoretical limit. In addition, there is no degeneration of the short-circuit behavior due to the relatively normal active cell mesa region. It can avoid the CIBL behavior induced by the very narrow conductive channel structure.

In this paper, an optimized trench power MOSFET is presented with three-dimensional (3D) TCAD simulation. A superior device’s performance can be achieved by using the partially widened split gate and the trench source. Simulation results indicate that the new structure can present a good static parameters and a better $R_{DS(\mathrm{ON})} {\times} \mathrm{Q}_{\mathrm{CD}}$ FOM compared with the conventional structure. Moreover, the proposed structure shows no SEB under the device’s rated breakdown voltage and the LET value of 1pC/$\mu$m $\left(100 \mathrm{MeV} \cdot \mathrm{cm}^{2} \cdot \mathrm{mg}^{-1}\right.$, approximately the maximum LET energy of gold or bismuth ion). That means, the proposed structure will not exhibit SEB under heavy ions. Therefore, the new structure demonstrates a good electrical parameter and a superior SEB SOA simultaneously.

A ultrahigh voltage p-channel 4H-SiC (Silicon Carbide) IGBT (Insulated Gate Bipolar Transistor) with CSL (Current Spreading Layer) structure is designed and optimized. The static and dynamic characteristics of the IGBT is improved by adjusting the CSL doping concentration $N_{CSL}$, the CSL thickness $D_{CSL}$ and the width of JFET region $L_{JFET}$. Through numerical simulations, when $N_{CSL}$ is (1×10)$^{16}$ cm$^{-3}$, $D_{CSL}$ is 1.5$\mu$ m and $L_{JFET}$ is 3$\mu$ m, the blocking voltage can reach 15.6kV, and the maximum of gate oxide field $E_{OX,max}$ is less than 3MV/cm. When the on-state current is 10A, the on-state voltage is only 8.2V. The turn-on time is about 120ns and the turn-off time is about 440ns at 5kV dc-link voltage.

Thread is an IPV6 based networking protocol for Internet-of-Things (IoT) that uses 6LoWPAN, IEEE 802.15.4 wireless at 2450MHz. OpenThread is an open source implementation of Thread Protocol released by Nest Labs that supports the Thread 1.1.1 specifications and currently used on Linux using command-line interface. This research focuses on building a Graphical User Interface (GUI) using TCL/TK for OpenThread. GUI helps researchers to manage and visualize Thread Protocol and its operations easier.

This paper presents an Internet-of-Things (IoT) architecture by integrating a Synology cloud server, edge computing systems, and physical networks. More specifically, we have established a physical network combining two subsystems 1) non-standardized Bluetooth Low Energy (BLE) Mesh network, and 2) a security monitoring system. The BLE Mesh system has one IoT host connected with three BLE devices, enabling to extend the communication distance by using one or two relays. The monitoring system consists of a Passive Infrared Sensor (PIR) and a webcam with multiple solutions for recognizing a human face. Two algorithms, Low Binary Pattern Histogram (LBPH) and Deep Metric Learning (DML), have been implemented and evaluated on different benchmarks. Experimental results show that the DML-based computation can reach 99.38% accuracy with almost 400 ms latency for recognizing a single face in frames of images. The future work will focus on testing the cloud service by integrating a Synology D218+ server, as well as improving the computation speed of facial recognition on pure hardware design on field-programmable gate array (FPGA). The aim of our work is to provide a robust IoT-Edge-Cloud system which can be deployed on the large-scale applications and processes data much faster compared to traditional cloud computing system due to the perks of parallel computing on FPGAs at the network edge.

Human has the ability of drawing inferences about other things from only one instance. Few-shot learning is aimed at imitating this generalized learning behavior of human beings, where the learning machine is expected to recognize novel categories not seen in the training set, given only a few training data for each novel category. In this paper, we enhance the Prototypical Network for few-shot learning tasks by introducing a domain alignment module, which takes into account the domain shifts existing between different categories. Compared to original Prototypical Network (PN), the most excellent model for few-shot learning at present, our proposed Domain Alignment Prototypical Network (DA-PN) is able to abate the distribution differences among the data of training and test classes, further optimizing the embedding space of prototype feature for each category and then boosting few-shot recognition. Comprehensive empirical evidence demonstrates that the proposed DA-PN can yield state-of-the-art few-shot recognition performance on the public benchmark dataset mini-ImageNet as well as a novel proposed few-shot dataset MNIST&CIFAR10.

A novel edge termination structure with Four-Region Multistep Field Limiting Rings (FRM-FLRs) is proposed and investigated for 10kV 4H-SiC IGBTs. The structure with FRM-FLRs terminal technology exhibits a good blocking performance with 30% less terminal area and 23% higher blocking voltage compared with the conventional Filed limiting Rings (Con-FLRs) structure. Moreover, the structure with FRM-FLRs can decrease the sensitivity of the interface charges, which greatly improves the reliability of the device at the off-state.

An optimized edge termination structure by increasing the coverage rate of metal plates at termination area is presented for 3.3kV IGBTs. This structure indicates a better stability after 168 hours High temperature reverse bias (HTRB) experiment. The electric potential contours and the surface electric field distribution with the variation of oxide charge are analyzed through numerical simulation. It is realized that the optimized termination structure is less sensitive with oxide charge. Thus the surface electric field is more uniform at each ring location to weaken the electric field concentration. Therefore, the breakdown voltage of the optimized structure is more stable than the conventional structure, which is in good agreement with the experiment results.

This paper reviews the security aspects of two Internet-of-Things (IoT) protocols, Z-Wave and Thread. Z-Wave is one of the oldest and most commercially successful IoT protocol, while Thread is one of the most recent protocols. As millions of IoT systems are installed for home and industrial automation, the security of these IoT systems is a concern. There is a potential for these IoT systems to be misused. This paper looks at security challenges for an IoT system. The security features of the IoT systems are spread across many parts of the IoT protocols. Paper discusses different attacks types on the IoT systems and the manners in which the protocols handle them. One of the most venerable times for an IoT system is when a new device is added to the IoT system. To avoid an intruder from getting access, the new device must be authenticated. Authentication of new network devices is discussed for both the protocols. IoT protocols are complex and the security aspects are also very complex, this paper should serve as a starting point to further study these and other IoT protocols.

Thread is a new IoT protocol released by Thread Group Inc. in 2015. More than 200 companies are currently members of the Thread Group Inc. and supporting this protocol. This paper provides an overview of the Thread protocol and discusses all the network layers. In the physical layer's overview we discuss the requirements for wireless communication. In the data link (MAC) layer we discuss the MAC link establishment (MLE) datagrams. We discuss use of 6LoWPAN addressing mode, which is a low power version of IPv6. In the network layer we discuss communication links and routing. In the transport layer we discuss the use of CoAP protocol and DTLS. In the security section we discuss setting up Commissioning agent and adding Thread devices to the Private Area Network.