Abstract: Neural networks have achieved state-of-the-art performance in solving many problems, including many applications in safety/security-critical systems. Researchers also discovered multiple security issues associated with neural networks. One of them is backdoor attacks, i.e., a neural network may be embedded with a backdoor such that a target output is almost always generated in the presence of a trigger. Existing defense approaches mostly focus on detecting whether a neural network is ‘backdoored’ based on heuristics, e.g., activation patterns. To the best of our knowledge, the only line of work which certifies the absence of backdoor is based on randomized smoothing, which is known to significantly reduce neural network performance. In this work, we propose an approach to verify whether a given neural network is free of backdoor with a certain level of success rate. Our approach integrates statistical sampling as well as abstract interpretation. The experiment results show that our approach effectively verifies the absence of backdoor or generates backdoor triggers.
Metaverse has recently attracted paramount attention due to its potential for future Internet. However, to fully realize such potential, Metaverse applications have to overcome various challenges such as massive resource demands, interoperability among applications, and security and privacy concerns. In this talk, we first discuss MetaChain, a novel blockchain-based framework to address emerging challenges for the development of Metaverse applications. In particular, by utilizing the smart contract mechanism, MetaChain can effectively manage and automate complex interactions among the Metaverse Service Provider (MSP) and the Metaverse users (MUs). We then proposed a novel framework, namely MetaSlicing, that can provide a highly effective and comprehensive solution in managing and allocating different types of resources for Metaverse applications. Specifically, by observing that Metaverse applications may have common functions, we first propose grouping applications into clusters, called MetaInstances. In a MetaInstance, common functions can be shared among applications. As such, the same resources can be used by multiple applications simultaneously, thereby enhancing resource utilization dramatically. To address the real-time characteristic and resource demand's dynamic and uncertainty in the Metaverse, we develop an effective framework based on the semi-Markov decision process and propose an intelligent admission control algorithm that can maximize resource utilization and enhance the Quality-of-Service for end-users.
After total thyroidectomy, Single-photon emission computed tomography (SPECT) is used to diagnose whether thyroid tissue remains in patients' bodies. Physicians visually diagnose the residual thyroid tissue in patient base on their expertise, so it is difficult for making a quick and accurate diagnostic. In present, Computer-aided Diagnosis systems (CAD) are becoming more widely in medical treatment, thyroid cancer is a potential field where CAD can improve diagnostic accuracy. This paper proposes a novel approach for diagnosing whether residual thyroid tissues remain in patient using thyroid SPECT scintigraphy by fine-tuning pre-trained Deep neural networks. Our proposed method achieved sensitivity of 84.85% and specificity is 89.11%, these results demonstrate that CAD is promising in diagnosing remaining thyroid tissue after total thyroidectomy.
The study was research in the bioinformatics field. The imagined signals are classified as the Support Vector Machine, K-Nearest Neighbor, and Ensemble Classifiers. A 5-channel device recorded the data, including four labels (thinking backward, thinking forward, thinking turn the left, and thinking turn the right). The data is optimized using z-score and maxmin normalization techniques and compared with data without normalization. The Stratified-Repeated cross-validation method was applied to split into training and testing data instead of the traditional data division technique. A key factor determining classifier performance is feature extraction. The time-frequency domain characteristics recommended by the Discrete Wavelet Transform method are five. The research examined 17 models (6 sub-model of Support Vector Machine and K-Nearest Neighbor classifiers, five Ensemble Classifiers). A model in the proposed Stratified-Repeated cross-validation Subspace Ensemble classifier with a classification result of 89.25%.
Our goal is to numerically solve optimization problems derived from a mechanical model of unilateral contact between solid bodies with rolling friction. The model is an optimization problem with a strictly convex quadratic objective function and a second-order cone of constraints that is not self-dual. The solver is an implementation of a primal-dual interior-point algorithm with the predictor-corrector scheme of Mehrotra extended to the second-order cone problem. We focused on analyzing the limits of numerical computation and proposed some treatments to achieve optimal solutions with ten significant digits of precision.
Colonoscopy is a common procedure for detecting and screening colorectal polyps, however, it poses a number of challenges because of the properties of images and the features of polyps inside, which can vary in shape, color, condition, and clear non-distinction with the surrounding context. Deep learning-based approaches have recently emerged, which considerably boost polyp segmentation procedures and gradually convert and replace out-of-date methods. Nevertheless, determining an effective measure remains a significant difficulty, which motivates us to propose GCA-Net for polyp segmentation. We first employ EfficientNetV2 as the backbone to extract improved salient features, then pass them via two modules, SE-PASPP and SE-RFB, to capture more contexture before feeding them into a dual-asymmetrical partial decoder to build a final resolution map. Finally, we develop a novel loss function based on the region-based strength of Dice loss and the geometrical constraints-based advantages of Active Contour with Elastica (ACE) loss. The suggested method's preeminence is demonstrated by comparing results to over state-of-the-art approaches in both learning and generalization assessment.
This paper proposes a new algorithm that reduces a large scale linear-time-invariant system having nonzero initial conditions into a smaller scale system. The estimation for the error between two outputs of the original and of the reduced systems is derived.
Software Defined Networking (SDN) with flexible control has a high potential and applicability in modern networks. SDN has outstanding advantages in actively and centrally controlling network functions such as dynamic-routing, load balancing, and preventing congestion for data flows to ensure the stability, fairness, and quality of the network system. For stability, response speed and fairness are required in modern networks. This paper proposes a new routing protocol called Fairness Enhanced Dynamic Routing Protocol (SDN-FERDP) to reduce the network congestion and optimize the load balancing of the network. The emulation result on Mininet shows that SDN-FERDP can effectively reduce network congestion and better balance the load of the network compared to the OSPF in the traditional network (simply refer to as OSPF) and Modified OSPF in the SDN network (refer to as SDN-OSPF).
Students who do not concentrate when studying will find it difficult to absorb the lesson well. Usually, in order for all students to focus on the lesson, the teacher during the lecture will have to observe the students and come up with solutions if the students are not paying attention. However, in the case of many students, following to detect students who have not paid attention to the lesson is a task that requires teachers to put in a lot of effort. In this article, we propose to use machine learning algorithms based on the MediaPipe library to analyze facial features and expressions, including eyes closed, yawning, not looking at the board, or absent, to determine if students have been distracted or not to build a system to assist teachers in detecting student lack of concentration when studying in Smart Desks (Student desks are designed based on embedded devices, with cameras and screens). When detecting that students are not paying attention while studying, the system will warn the teacher so that the teacher can provide solutions. We tested the algorithm on a Jetson Nano embedded device with configuration [Quad-Core 64-bit ARM, 128-bit GPU CUDA, 4GB RAM] and obtained FPS: 8 ~ 18, accuracy achieved from 89 ~ 97% in lighting conditions from 300–400 lux.
Small Size League (SSL) is a division of the traditional RoboCup, founded to promote research in robots and AI. A fast and accurate real-time object detection model is essential for RoboCup SSL soccer robots, serving the design and development of competitive strategies. Specific state-of-the-art object detection methods have reported inference speed up to 94 FPS on the SSL open-source benchmark dataset, but only at intermediate accuracy. Considering the advancement in deep learning methods for feature extraction and object detection, in this paper, we conducted surveys and experiments on one-stage object detection methods provided in the MMDetection framework on the dataset for RoboCup SSL. YOLOX-tiny model achieved 58.60% AP, which is significantly higher than baseline methods, while maintaining an acceptable inference speed of 37 Frames Per Second (FPS). Other state-of-the-art one-stage methods have achieved very high performance, up to 74,10% Average Precision (AP). However, certain methods did not meet the minimum inference speed requirement of real-time object detection.
Water pollution become an serious problem in nowadays. The water can be polluted by many factors, including chemicals, trash, bacteria, and parasites. Different with rest pollutants, which need complex experiment to determine the pollution level, trash can be easily to detect by human eye. However, this work may take a numerous cost while monitoring on large area or for long time, which can also easily increase the errors. Therefore, an effective solution need to be explored to reduce not only the cost but also the errors. This paper proposes an efficient model for automatically detection of floating trash based on YOLOv5s. By using a lightweight architecture, our model give a comparative performances with the original model on different benchmarks, which prove the effectiveness of the proposed method. So, our method could be applied to any monitoring or detecting systems with low cost.
Today, customer-centric business models are more focused. In addition, businesses increasingly own more data about customers and user behavior, especially from social networking platforms. However, the effective exploitation and use of this amount of data is still very limited. Based on those objective motivations, this research proposes a model as a foundation for analyzing and understanding customers to help businesses transform and transform this data into forms of knowledge to support the business, make intelligent business decisions. A data set of 90,513 reviews in Vietnamese from Agoda e-commerce websites in the tourism sector is collected and experimented. In experimental phase, raw corpus was cleaned by text preprocessing, then, the research will use the Net Promoter Score (NPS) to separate the data into three data sets according to three groups of Detractors, Passives and Promoters. From that, Latent Dirichlet Allocation (LDA) was built to extract the hidden topics in the input dataset. The experimental results show that the extracted topics and their features from corpus of customer reviews are reliable. The novelty in this research is to identify the topics that customers are interested in and complaining about.
In this paper, we have solved the four key problems of autonomous ping-pong ball collecting robots: ball detection, distance estimation, path planning, and avoiding objects in the path. The proposed path planning and obstacle avoidance method are the most effective compared to current practices in simulation. Practical experiments also give promising results.
This paper presents a study on the problem of information extraction and its application in building an information extraction system for administrative documents. The proposed end-to-end system contains three significant modules, including Text detection (TD), Optical character recognition (OCR), and Information extraction (IE). We developed the IE module by us based on two platforms, GraphSAGE and GATs. We have made many changes and improvements, such as redesigning graph modeling and node representation to match the goals and problems posed. We also elaborately studied to establish a complete information extraction system and dived into the information extraction module instead of all modules in the system. Besides that, we also built and evaluated our dataset of Vietnamese Administrative Documents Images (VADI2021).
Nowadays, in the age of big data, cloud computing, and the internet of things, there has been the ubiquity of open data for people to conveniently approach, use, and mine. However, most of the valuable data is data relating to personal sensitive information such as data about diseases or salaries. This data should be protected by some policies to conceal the relationship between sensitive data and people. As of late, there are many approaches and techniques for preserving data privacy, of which k-anonymity is the most popular. Nevertheless, most of the k-anonymity algorithms are too general and do not concentrate on any concrete data mining technique, so the data utility does not remain high. In this work, we introduce a k-anonymity algorithm based on tuple member migration between tuple groups to achieve k-anonymity while preserving data quality for a data mining algorithm, i.e., association rule mining because it is one of the most popular data mining techniques which discovers the association of items or itemsets in a dataset. The algorithm was evaluated on the adult dataset to assess the performance as well as the data utility.
This paper discusses the optimal aerodynamic shape design for the vehicle model participating in the Shell Eco-Marathon. The design of the vehicle's body is developed based on the chassis structure and regulations of the competition. Simultaneously, the NACA 2421 Airfoil profile is used to construct the vehicle body shape to reduce the aerodynamic drag. Three series of numerical simulations were conducted to determine the optimal values of the diffuser angle and the body side profile for the minimum aerodynamic drag. The first simulation series was performed on a three-dimensional vehicle model with different diffuser angles to find the angle that yields the smallest aerodynamic drag value. In the second series of simulations, the optimization focused on the body side profile of the vehicle model with different heights of the NACA 2421 profile sections of 0.5H, 0.6H, and 0.7H, in which H is the total height of the upper side of the NACA 2421 airfoil's profile, while the optimal diffuser angle is fixed. The last simulation series investigated the aerodynamic drag on the vehicle model with the optimal diffuser angle and body side shape at different velocities. The results show that the model with a diffuser angle of 15° and a body side profile of the NACA 2421 profile section with the height of 0.6H produces the minimum aerodynamic drag. The optimal drag coefficients of the vehicle model vary in the range of 0.150 to 0.129 along with the velocities of 20 to 50 km/h, respectively. The study showed that applying the NACA 2421 profile to the vehicle profile design for reducing fuel consumption has obtained significant achievements.
Voice activity detection in the wild is considered to be challenging work, especially when applied to the Vietnamese language as many proposed approaches are not extensive enough. In this paper, we aim to solve this problem by using MarbleNet, a model built on top of previous successful applications of using ID CNNs to solve conventional problems. We compiled a dataset, a combination of the VIVOS dataset for speech labelling and audios collected from Freesound.org for background noise. We present the performance of MarbleNet on the chosen dataset and perform experiments that compare the performance of MarbleNet and two other CNN-based architectures to measure the efficiency of our solution. Experiments show that MarbleNet, with a smaller size, can outperform other CNN-based models in clean and many noisy environments.
In this paper, we investigate a wireless body area network (WBAN) with multiple users and multiple access points (AP). A user has multiple sensors which are delivered around the body. In a conventional WBAN, a sensor transmits its own data to the selected AP with full transmit power. Consequently, it affects the transmission of other sensors, resulting in a significant degradation in system performance. In this study, the authors discuss and propose a power control algorithm to increase the spectral efficiency of sensors under the condition of constrained transmitting power. According to simulation results, most sensors achieve higher spectral efficiency with the proposed power control approach than without power control. Furthermore, it indicates that the proposed power control algorithm significantly improves the WBAN system's throughput in both the uplink and downlink communications.
This paper is concerned with secure communication in multiuser full-duplex (FD) communication systems with simultaneous wireless information and power transfer (SWIPT). In the considered system, an FD base station (BS) equipped with multiple antennas communicate with both uplink users (ULUs) and downlink users (DLUs) at the same time and in the same frequency while the set of energy harvesters can scavenge energy from ambient radio frequency signals. To keep the uplink and downlink information signals confidential from the energy harvesters which can be potential eavesdroppers and to optimize the harvested energy, we aim at designing the precoders and the transmit artificial noise covariance at the BS and the transmit precoders at ULUs to maximize the harvested energy subject to the constraints of the required minimum quality of service (QoS) and transmit power at the transmitters. Due to high nonconvexity and nonlinearity of the formulated design problem, we develop an iterative algorithm based on successive convex programming (SCP) by seeking appropriate surrogate functions to recast the original optimization problems into convex optimization ones. Then, the simulation results are conducted to evaluate the system performance in terms of the sum harvested energy for different system setups.
Trigonometric operations have a wide range of use in communication, signal processing, and especially in computer science. Many methods exist to implement these functions on hardware, but complicated algorithms lead to high hardware consumption and latency. This paper presents a design to perform trigonometric calculations on FPGA that can be processed in parallel to reduce latency. We propose loop-optimized Radix-4 CORDIC algorithm for hardware implementation. This algorithm uses only three iterations to compute high accuracy trigonometric values. Besides, we apply multiply-less hardware architecture for the design, which consists of three basic operations: adders, subtractors, and bit shifters. The design is implemented on the Zynq™-7000 AP SoC kit XC7Z020-CLG484-100 device. The performance results show that the output returns values with absolute error lower than 0.005 after three clock cycles.
