Search

Reinforcement learning (RL) is rapidly used in safety-centric applications. However, many studies focus on generating optimal policy that achieves maximum rewards. While maximum rewards are beneficial, safety constraints and non-functional requirements must also be considered in safety-centric applications to avoid dangerous situations. For example, in the case of food delivery robots in restaurants, RL should be used not only to find optimal policy that response to all customer requests through maximum rewards but also to consider safety constraints such as collision avoidance and non-functional requirements such as battery saving. In this paper, we investigated the fulfillment of safety constraints and non-functional requirements of learning models generated through RL with quantitative model checking. We experimented with various time steps and learning rates required for RL, targeting restaurant delivery robots. The functional requirement of these robots is to process all customer order requests, and the non-functional requirements are the number of steps and battery consumption to complete the task. Safety constraints include the amount of collision and the probability of collision. Through these experiments, we made three important findings. First, learning models that obtain maximum rewards may have a low degree of achievement of non-functional requirements and safety constraints. Second, as safety constraints are met, the degree of achievement of non-functional requirements may be low. Third, even if the maximum reward is not obtained, sacrificing non-functional requirements can maximize the achievement of safety constraints. These results show that learning models generated through RL can trade off rewards to achieve safety constraints. In conclusion, our work can contribute to selecting suitable hyperparameters and optimal learning models during RL.

Microgrids are fundamental elements in modern energy systems. Among the various microgrid components, the Energy Storage System (ESS) plays a pivotal role in ensuring system reliability, but its high cost and inevitable degradation over time pose significant challenges. Many current studies overlook the impact of ESS degradation on operational optimization, potentially leading to cost-ineffective systems. To address this gap, we introduce a quadratic ESS degradation model that captures intricate battery dynamics, such as State of Charge (SoC) and Depth of Discharge (DoD), using Markovian properties. Based on this model, we propose an optimal energy management framework for DC microgrids using Quadratic Programming (QP). The objective is to minimize the combined costs of degradation and electricity, considering the Time-of-Use (ToU) tariff while adhering to ESS constraints. This financially focused approach provides a pragmatic and economically aligned optimization strategy. Testing across various State of Health (SoH) scenarios demonstrates that our proposed model reduces total operational costs by 3-18%. This research advances microgrid optimization techniques and offers practical insights to enhance efficiency and economic resilience in real-world scenarios.

The sixth generation (6G) is the next generation of wireless communication technology, is not limited to cellular networks but can be used to provide better services in all areas of wireless communication, including vehicles, drones, and smart homes. However, these advancements in 6G technology require further security considerations. In earlier networks, authentication schemes were designed to rely on a secure channel between the core network and access points. The 6G network is an open, distributed network that integrates multiple domains and entities and cannot guarantee secure channels in the network. The 6G network requires a new security authentication scheme that applies the zero-trust model. Therefore, this study proposes a zero-trust authentication scheme with access control for 6G-enabled Internet of Things environments. The scheme uses blockchain technology for mutual authentication in a distributed environment and lightweight attribute-based encryption to ensure dynamic access control and network efficiency. This study compares the proposed authentication method with existing methods and demonstrates that this scheme has better performance and security. To the best of our knowledge, this paper is the first to propose a specific authentication protocol with access control considering the zero-trust model in a 6G environment.

This paper proposes a reliable data dissemination framework for edge networks, leveraging network coding combined with low-rank approximation. We consider an edge network that consists of a server and power-limited mobile devices, where the data is broadcasted by the server. In such networks, broadcasted data may be lost due to poor channel conditions or the interference caused by the mobility of edge mobile devices, particularly without a retransmission mechanism. This can cause application errors in edge devices, lower the Quality of Service (QoS), and compromise network stability. To overcome these challenges, we propose a framework for reliable edge networks in broadcasting without retransmissions. The edge network reliability can be achieved by the approximate decoding of broadcasted data. In the proposed framework, the edge server employs matrix factorization to encode data with principal components, ensuring a lower decoding error rate even with potential packet losses. Furthermore, the proposed framework can shift the computational complexity from mobile edge devices to the edge server using the low-rank approximation at the decoding stage, effectively mitigating power limitations on mobile devices. Through theoretical analysis, we demonstrate that the proposed algorithm outperforms typical broadcasting in terms of decoding accuracy, and present an upper bound error rate for the proposed algorithm. The simulation results confirm that the proposed algorithm outperforms other state-of-the-art algorithms in terms of decoding accuracy, time delay, and complexity.

Deep neural networks (DNNs) provide excellent performance in image recognition, speech recognition, video recognition, and pattern analysis. However, they are vulnerable to adversarial example attacks. An adversarial example, which is input to which a little bit of noise has been strategically added, appears normal to the human eye but will be misrecognized by the DNN. In this paper, we propose AdvGuard, a method for resisting adversarial example attacks. This defense method prevents the generation of adversarial examples by constructing a robust DNN that provides random confidence values. This method does not require training of adversarial examples, use of other processing modules, or the ability to perform input data filtering. In addition, a DNN constructed using the proposed scheme can defend against adversarial examples while maintaining its accuracy on the original samples. In the experimental evaluation, MNIST and CIFAR10 were used as datasets, and TensorFlow was used as a machine learning library. The results show that a DNN constructed using the proposed method can correctly classify adversarial examples with 100% and 99.5% accuracy on MNIST and CIFAR10, respectively.

This study proposes a novel motion planning strategy to address localization uncertainty in autonomous buses. Conventional motion planning algorithms utilize information from high-definition (HD) maps to overcome the limited detection range of on-board sensors. However, this information contains uncertainty due to the utilization of estimated localization results during reference frame transformation. The wider dimensions of autonomous buses, compared to regular vehicles, amplify the potential dangers associated with localization uncertainty. Therefore, this research focuses on enhancing motion planning for autonomous buses by effectively addressing localization uncertainty. The investigation of manual driving data from autonomous buses highlights the need to handle three issues: heading bias, lateral position error, and longitudinal position error. Firstly, the heading bias was dealt with by implementing an Offset-free Model Predictive Control (OF-MPC) with a Moving Horizon Estimation (MHE) scheme for lateral motion planning. Secondly, the lateral position error was handled by incorporating a drivable corridor to determine the desired path. Lastly, the longitudinal position error was resolved by implementing a chance-constrained MPC for longitudinal motion planning. The proposed approach showed noticeable enhancement in path tracking performance while still securing ride comfort in lateral motion, collision safety, and prevention of stop-line violations. We evaluated the feasibility of the proposed approach through vehicle tests on a test track, and its applicability was further confirmed through fully autonomous driving tests on actual urban bus-only lanes.

As Generative Adversarial Networks advance, deepfakes have become increasingly realistic, thereby escalating societal, economic, and political threats. In confronting these heightened risks, the research community has identified two promising defensive strategies: proactive deepfake disruption and reactive deepfake detection. Typically, proactive and reactive defenses coexist, each addressing the shortcomings of the other. However, this paper brings to the fore a critical yet overlooked issue associated with the simultaneous deployment of these deepfake countermeasures. Genuine images gathered from the Internet, already imbued with disrupting perturbations, can lead to data poisoning in the training datasets of deepfake detection models, thereby severely affecting detection accuracy. We propose an improved training framework to address this problem in deepfake detection models. Our approach involves purifying the disrupting perturbations in disruptive images using a backward process of the denoising diffusion probabilistic model (DDPM). Images purified using our DDPM-based technique closely mimic the original, unperturbed images, thereby enabling the successful generation of deepfake images for training purposes. Moreover, our purification process outperforms DiffPure, a prominent adversarial purification method, in terms of speed. While conventional defensive techniques struggle to preserve detection accuracy in the face of a poisoned training dataset, our framework markedly reduces this accuracy drop, thus achieving superior performance across a range of detection models. Our experiments demonstrate that deepfake detection models trained using our framework exhibit an increase in detection accuracy ranging from 11.24%p to 45.72%p when compared to models trained with the DiffPure method. Our implementation is available at https://github.com/seclab-yonsei/Anti-disrupt.

Voltage source converter (VSC)-based high-voltage direct-current (HVDC) systems are being increasingly considered to support grid frequency regulation (FR) by using fast responses to reference signals. This paper proposes a control strategy for the optimal real-time FR of multi-terminal (MT) HVDC grids to reduce grid frequency variations. The novelty of the proposed strategy is that it realizes coordinated power control (CPC) of MT-HVDC system to consider the malfunctions of the HVDC converters and communication time delays between the malfunctioning and compensated HVDC converters. Accordingly, a state-space model of MT-HVDC grids is derived considering the DC-link voltage and active power controls at the rectifier and inverter, respectively, which are combined with inertia emulation (IE) and droop controllers. Based on the developed model for MT-HVDC grids, optimal secondary frequency control (SFC) is achieved using a linear quadratic Gaussian (LQG) controller. Moreover, CPC is integrated into the LQG to compensate for the inaccurate output power owing to malfunctions of the HVDC converter. Simulation case studies are performed to verify that the proposed strategy can alleviate real-time frequency fluctuations of inverter-side grids more effectively than conventional strategies.

We recently established an impedance standard for the D-band, one of the 6G candidate frequencies. However, primary standards are difficult to use for routine calibration because they require multiple impedance standards and a lot of time to calibrate the vector network analyzer (VNA). Therefore, a transfer standard is needed to efficiently apply the calibration value and to propagate the uncertainty of the primary standard to a device under test (DUT). In this paper, we describe a design method for a transfer standard with small uncertainty even when an arbitrary DUT is measured. This is achieved by propagating the uncertainty of the primary standard to the transfer standard and then propagating it again to the uncertainty of the DUT. We developed a calibration kit that has low uncertainty over a wide frequency band, from 110 GHz to 170 GHz, and consists of waveguide offset shorts for ease of production. We also propose a method to minimize DUT uncertainty and a method to minimize “phase distance” to find the optimal length of the offset short. When using three offset shorts, an uncertainty similar to that of the short-open-load-thru (SOLT) calibration kit was obtained, and when using four offset shorts, an uncertainty comparable to the primary standard was obtained. Lastly, this paper examines the repeatability of measurements and reproducibility during the production process.

In this study, we propose an automated system for measuring the size of strawberries and predicting their weight using AI technology. The system combines computer vision techniques with LiDAR sensor data to accurately estimate the dimensions of strawberries and infer their weight. By integrating deep learning models, such as HRNet for keypoint detection, and leveraging the capabilities of LiDAR sensors, we minimize human intervention and achieve precise size measurement. The relative errors for the width and height of the strawberries are 3.71% and 5.42%, respectively, with the width exhibiting a lower error rate. The standard deviation for the width and height of the strawberries are 0.19% and 0.24%, this indicates that the individual strawberries had very low error rates in terms of their measurements for the width and height. Weight prediction was performed through regression analysis with width and height estimation. Experimental results demonstrate that our approach enables accurate weight prediction with a relative error of 10.3%. This automated technology holds great potential for strawberry harvesting and classification tasks, facilitating the automation of these processes.

Advanced technologies, driven by extensive data analysis, support the concept of intelligent cities, which aim to enhance the quality of people’s lives, minimize the consumption of energy, reduce pollution, and promote economic growth. The transportation network is a crucial component of this vision in urbanized cities. However, a massive increase in road traffic poses a significant challenge to achieving this vision. Developing an intelligent transportation system requires accurately predicting the traffic speed. This paper proposes a novel deep stacking-based Ensemble model with a two-layer structure to address the problem of forecasting traffic speed in urbanized transportation networks to solve traffic congestion problems. Firstly, advanced machine learning such as eXtreme Gradient Boosting(XGB), Random Forest(RF), and Extra Tree(ET) as base learners are used to predict short-term traffic speed. In the next phase, the Multi-Layer Perceptron (MLP) as a meta-learner technique, employing various combinations of the aforementioned approaches is used to enhance the accuracy of traffic speed prediction. The proposed stacking-based approach has the capability to analyze, extract, and aggregate various features from primary traffic speed data in order to generate more refined and accurate forecasts. This study used a publicly available dataset of Floating Cars Data collected from real transportation networks for evaluation. Mutual information regression is used as a feature selection technique to obtain the features from the dataset for the training of these models. The performance results are compared with state-of-the-art traffic prediction models. Results show that the proposed stacking-based ensemble strategy outperforms conventional approaches by a large margin such as HA, KNN, SVR, DT, T-GCN, and A3TGCN models. The results demonstrate a notable reduction of 9.71% in RMSE and 15.4% in MAE, indicating enhanced accuracy. Furthermore, our approach achieved a substantial improvement of 13.80% in $\mathbb {R}^{2}$ and 11.64% in EV for the 15-minute prediction horizon.

The DC-DC converters are essential in power electronics as they maintain a stable output voltage even when there are changes in input voltage and load current. This study introduces an advanced Proportional-Derivative (PD) compensator for buck converters. The compensator enhances stability and transient responsiveness by employing a unique modulation technique that has not been previously applied in this context. The proposed method entails applying an input of 28 volts, which yields an output amplification of 15 volts, even in the presence of interference. The small signal transfer function of the buck converter is meticulously derived, considering the converter’s dynamic behavior to achieve exceptional results. The transfer function comprehensively explains the intricate relationship between the input and output voltages, providing the theoretical basis for our distinctive control approach. The MATLAB code accurately generates the Bode diagram of the buck converter by employing the small signal transfer function. The graph illustrates the frequency response of the converter, a crucial factor in enhancing the stability and quality of the output voltage. The proposed research is substantiated by mathematical data shown through several simulated figures, distinguishing it from conventional methodologies. The implemented research achieves the maximum efficiency exceeding 95% with a minimum ripple factor.

There has been a surge in research focused on the analysis of vital signs using remote photoplethysmography (rPPG) sensors, as opposed to traditional photoplethysmography (PPG) methods. Unlike PPG, rPPG imposes no spatial constraints and employs a straightforward measurement technique, making it increasingly prevalent. Its integration into image processing, harnesses the remarkable advances in artificial-intelligence technology, achieving accuracy that is comparable to that of traditional PPG sensors. In prior studies, obtaining vital signs often necessitated an unnecessary and procedural fixation of facial positions within frames to enhance predictive accuracy. Despite such fixation, achieving notably high accuracy remained elusive. Here, we introduce a simple yet robust approach utilizing videos captured by an rPPG sensor, ensuring both high accuracy and resilience to noise. We propose a convolutional neural network model meticulously designed to resist interference from noise data that may arise in the initial stages, coupled with effective preprocessing techniques to attain superior predictive accuracy. Data extracted by a facial extractor undergoes preprocessing via normalization. Leveraging the Temporal Shift Module (TSM), this normalization efficiently captures temporal relationships without incurring additional computational overhead. Mitigating noise signal interference from non-facial data through the use of multiple attention masks and augmenting prediction accuracy via skip connections. Moreover, we compile a specially tailored dataset for pulse rate and breath rate data, catering specifically to the East Asian population. The proposed process demonstrates outstanding performance in predicting both pulse rate and breath rate.

While Smart Grid offers high efficiency in power delivery, it is susceptible to cyberattacks because of vulnerabilities in the information and communication technologies. Network segregation lowers threats by limiting their consequences within segregated network. Network segregation can be achieved either logically or physically. Logical segregation relies on firewalls to filter and manage network traffic. Whereas physical segregation employs methods such as air gaps or data diodes, which provide heightened security by necessitating physical access for a breach. Although air gaps entirely isolate domains from communication, data diodes allow only unidirectional data flow. Effective communication regulation between domains is emphasized owing to its restricted nature, leading to the development of cross-domain solutions (CDS). Certain types of CDS facilitate bidirectional communication by combining two data diodes. The issue lies in the inability of current CDS solutions to consider application-level protocol intricacies. The Modbus protocol is a representative example. To ensure secure communication, a CDS must match the incoming response packets with outgoing request packets, which requires the extraction and correlation of state variables. However, the current CDS, next-gen firewalls, and intrusion prevention systems lack this capability. Thus, this study proposed a next-gen CDS architecture capable of stateful correlation of outgoing and incoming application-layer packets. The proposed method extracts user-defined state variables from outgoing traffic and evaluates incoming packets based on rulesets. A prototype based on this method exhibits superior filtering accuracy compared to traditional CDS, despite a 51.08% increase in processing delay, thereby demonstrating its potential in enhancing network security.

This study addresses the loss minimization and voltage stability analysis for low-frequency transmission lines connected to wind farms, pumped storage Hydropower (PSH) and Gas Turbine plant. The losses and the voltage stability index connected with renewable energy systems are heavily dependent on the operating frequency and create uncertainties in terms of voltage and frequency. Therefore, optimal frequency selection for LF-HVac lines also become important for the system. The voltage stability has been improved by STATCOM (Static Synchronous Compensator) for given system placed at different buses to maintain stability of system. A simplified two-port equivalent model for transmission lines based on solutions to Telegrapher’ method has been proposed for reducing system complexity and losses. The proposed method is tested with different existing algorithms and implemented in IEEE 39-Bus system and for scalability compared with IEEE 57-Bus system. The results obtained in MATLAB environment in Simulation lab which indicate that the proposed LF estimation method with the modified metaheuristic algorithm gives optimal frequency for low-frequency HVac systems.

Location-based pairwise key predistribution protocols have been developed as promising solutions for addressing security concerns in wireless sensor networks. These methods efficiently distribute symmetric keys among sensor nodes based on their locations, thereby enhancing security. However, prior designs often assumed ideal node locations without giving sufficient consideration to possible deployment errors. Such an error, causing a node to move out of its intended grid cell location, can significantly impact network connectivity. In real-world scenarios, these errors may arise due to 3D terrain features, such as slopes and other physical environmental factors. Unfortunately, these factors have often been overlooked by existing location-based schemes, indicating the necessity for revisiting these methods. In this paper, we bridge this gap by simulating node deployments across various 3D terrains while considering physical factors. We scrutinize the correlation between physical environmental factors and deployment errors and evaluate the performance of both computation-based and storage-based protocols. Furthermore, we identify conditions that minimize deployment errors, thereby enhancing the performance of location-based pairwise key predistribution protocols. Our experimental results show that terrain features and external physical factors significantly influence the performance of location-based protocols, and their consideration is necessary.

Smart cars have become more intelligent and technologically complex; however, consumers have raised concerns about their security because of hacking and technical safety. Although numerous studies have focused on the technological issues of smart cars, there is a need for new perspectives and discussions to examine drivers’ perceptions of smart car security and safety. Understanding drivers’ security perceptions and behaviors in a smart car environment is crucial for enhancing smart car security and driving industrial growth. Therefore, this study analyses the factors affecting drivers’ willingness to pay (WTP) for smart car security based on protection motivation theory and explores the relationship between protective motivations and post-behavior. This study analyses how drivers’ security vulnerabilities and severity perceptions in a smart car environment affect their WTP for security software. In addition, the comparative analysis discusses how security perceptions differ between traditional PC and smart car environments. The results of this study demonstrate that drivers are sensitive to smart car security and are willing to pay to strengthen security. Moreover, the results of this study suggest that the role of security solution companies is important for strengthening smart car security and expanding the market. This study has academic significance because it is an early study discussing driver behavior related to smart car security software on a theoretical basis. It is also significant because it provides practical implications for smart car security market growth and guides the formulation of effective security policies.

Recent advances in quantum-computing technology have threatened the security of classical cryptographic algorithms. This initiated research on Post-Quantum Cryptography (PQC), and the National Institute of Standards and Technology (NIST) PQC standardization is in progress. Coping with the current situation in which the security of existing cryptographic algorithms is already in question and that of new cryptographic algorithms is not yet certain, there has been active research on hybrid schemes combining two algorithms such that the security of the combined scheme is based on both underlying algorithms. For digital signatures, a naive solution for a hybrid scheme is to simply concatenate a classical signature and a quantum-resistant signature. In this paper, however, we propose a compact hybrid signature construction method that combines two randomized signatures such that the size of the combined signature is shorter than that of naive concatenation. Our construction allows for selective verification, which provides backward compatibility and conformance with existing regulations. We demonstrate the feasibility of the proposed method by combining ECDSA P-256 and Falcon-512, which are representative classical and post-quantum signature schemes, respectively. We prove that the combined signature is existentially unforgeable against an adaptive chosen-message attack, even if one of the underlying signature schemes is completely broken and only the other one remains secure. Through experiments on a desktop PC and Raspberry Pi 3 Model B, we verify that the proposed method effectively reduces the combined signature size with negligible computational overhead. Our experimental results demonstrate the proposed method is also applicable to PQC-PQC combinations.

This paper presents the novel control method of the Four-Switch Buck-Boost (FSBB) converter for Super Capacitor Pre-Charger (SCPC) in hydrogen electric vehicles. The method is proposed to improve voltage regulation characteristic during super capacitor pre-charging, and is able to maintain continuity of duty ratio and voltage conversion ratio. The novel control scheme uses two independent duty ratio, and one varies while the other is fixed in whole operation. The duty ratio of changing the value changes at the transition point. Even though operation mode changes for achieving unity voltage gain, the additional switch state diminishes voltage fluctuation at transition point compared to conventional method. Also, the size of passive components can be reduced. The feasibility of the proposed control method is substantiated through simulation and experiment with 10kW prototype.

Voice phishing is the criminal act of tricking others to transfer funds or to seek financial gain based on personal information obtained illegally. The importance of this crime is recognized worldwide, and technical solutions have been proposed to reduce the increasing damage. In this paper, we propose a process for voice phishing detection in Korean by applying named entity recognition (NER) with Key Tags and Sentence-level N-gram. From the perspective of humans, we collect financial counseling texts as non-phishing dataset since the victim confuses voice phishing with them. We carefully select Key Tags that can be meaningful for distinguishing voice phishing and financial counseling texts and combine sentence bundles to effectively detect voice phishing. The experimental results, using ten types of machine learning models, showed that maintained results when generalizing information by Key Tags and improved results when combining text bundles. We hope that the proposed process can be effectively applied to other criminal scenarios in the future.

Drones have become integral in modern warfare, and their evolution towards greater autonomy is inevitable. This study explores the trajectory towards intelligent, minimally human-dependent military drones, detailing the necessary technological advancements. We simulate drone reconnaissance operations to identify and analyze emergent challenges. This study delves into various technologies critical for enhancing drone intelligence, with a focus on reinforcement learning based on object detection, and presents practical implementation examples to illustrate these advancements. Our findings confirm the significant potential for enhancing the intelligence of military drones, paving the way for more autonomous and effective combat solutions.

Currently, in the semiconductor market, reliability response to Electro Static Discharge (ESD) situations is being discussed as an alternative to internal Integrated Circuit (IC) destruction. In this paper, the ESD protection circuit design integrated into the Low Drop Out (LDO) regulator presents an alternative to the reliability of the internal IC. Additionally, the LDO regulator proposed in this paper is designed to minimize output voltage changes by significantly improving gain using a high gain buffer structure. Undershoot and overshoot were effectively regulated by utilizing a high gain buffer. Under the conditions of an input voltage range of 3.3–4.5 V, a maximum load current of 400 mA, and an output voltage of 3 V, the output of the proposed LDO regulator with a high gain buffer structure kept at the undershoot and overshoot voltages of 32 mV and 36 mV, respectively. In addition, the ESD protection circuit used so far was designed with only a single diode, but the proposed ESD protection circuit is a new structure designed based on Silicon Controlled Rectifier (SCR). The proposed ESD protection circuit has an incomparable area efficiency compared to diodes.

Issues are evolving requirements in software engineering. They are the main factors that increase the cost of software evolution. To help developers manage issues, GitHub provides issue labeling mechanisms in issue management systems. However, manually labeling issue reports still requires considerable developer workload. To ease developers’ burden, researchers have proposed automatically classifying issue reports. To improve the classification accuracy, researchers adopted deep learning techniques and pretrained models. However, pretrained models in the general domain such as RoBERTa have limitations in understanding the contexts of software engineering tasks. In this paper, we create a pretrained model, IssueBERT, with issue data to understand whether a domain-specific pretrained model could improve the accuracy of issue report classification. We also adopt and explore several pretrained models in the software engineering domain, namely, CodeBERT, BERTOverflow, and seBERT. We conduct a comparative experiment on these pretrained models to evaluate their performance in classifying issue reports. Our comparison results show that IssueBERT outperforms the other pretrained models. Noticeably, IssueBERT yields an average F1 score that is 1.74% higher than that of seBERT and 3.61% higher than that of RoBERTa, even though IssueBERT was pretrained with much less data than seBERT and RoBERTa.

Personality detection plays a pivotal role in social interactions, machine learning (ML), and natural language processing (NLP). Its goal is to discern an individual’s traits from their behavior and expressions. The prevalence of text-based communication has sparked interest in inferring personality from written content. However, challenges persist in accurately interpreting traits like the Big-Five or Myers-Briggs Type Indicator. These challenges stem from the reliance on self-reported surveys for labeling, which introduces uncertainties as individual assessments may not consistently align with their actual personality. In this paper, we propose novel curriculum strategies that employ class-based term frequency-inverse document frequency (c-TF-IDF) to enhance personality detection performance. By leveraging a curriculum approach that mirrors human learning progression, starting from simpler tasks and moving toward more complex ones, these strategies aim to train models on progressively challenging scenarios. Our experimental results demonstrate that these proposed curriculum-based strategies improve the accuracy of personality detection compared to previously suggested methods. This study contributes to advance understanding of text-based cues for personality inference. It has the potential to enrich various fields, including human-computer interaction, personalized recommendations, and targeted marketing.

In holography, the resolution of the hologram significantly impacts both display size and angle-of-view, yet achieving high-resolution holograms presents formidable challenges, whether in capturing real-world holograms or in the computational demands of Computer-Generated Holography. To overcome this challenge, we introduce an innovative Hologram-to-Hologram Super-Resolution network (H2HSR) powered by deep learning. Our encoder-decoder architecture, featuring a novel up-sampling block in the decoder, is adaptable to diverse backbone networks. Employing two critical loss functions, data fidelity and perceptual loss, we guide H2HSR to attain pixel-wise accuracy and perceptual quality. Rigorous evaluations, using the MIT-CGH-4K dataset, demonstrate H2HSR’s consistent superiority over conventional interpolation methods and a prior GAN-based approach. Particularly, in conjunction with the SwinIR encoder, H2HSR achieves a remarkable 8.46% PSNR enhancement and a 9.30% SSIM increase compared to the previous GAN-based method. Also, we found that our H2HSR shows more stable reconstruction quality across varying focal distances. These results demonstrate the robustness and effectiveness of our H2HSR in the context of hologram super-resolution.

As the importance of trajectory analysis arises in video surveillance, it becomes crucial to define the dissimilarity measure between two trajectories. Although the Hausdorff distance can be considered as a viable candidate for the measure, it is challenging to deal with noise present in trajectories since the Hausdorff distance is susceptible to noise so that even a single noise point may significantly distort the distance computation. In this paper, we propose a novel approach to alleviate the influence of inherent noise by setting noise-like points apart from ordinary points with a novel spatial tree structure during trajectory distance computation, without additional noise detection processes. In particular, we present $\mathrm {R}^{on}$ -tree, an extension of the existing spatial tree structure, that seamlessly finds permanent noise-like points, which are considered to have a low possibility of being ordinary points, and then keeps them in a separate auxiliary R-tree, without any separate process of disclosing noise-like points. We exploit $\mathrm {R}^{on}$ -tree to compute the noise-tolerant trajectory distance by modifying an existing algorithm for the Hausdorff distance. We also build an algorithm for noise-tolerant trajectory search to ensure accurate and high-quality search results even with noisy trajectories. The empirical results show that in all cases, our proposed approach yields the distance closest to the true one than any other competitor. The effectiveness of our approach is further examined by applying our noise-tolerant trajectory search to a real video surveillance dataset.

In the emergency department, accurately identifying patients who will require intensive care unit admission is crucial for effective care and resource allocation. Existing methods, such as triage judgments based on Emergency Severity Index, may not fully capture the dynamic nature of patient conditions, potentially overlooking critical risks. Therefore, it is essential to detect sequential deterioration in patients with initially underestimated triage scores. This study addresses the challenge of identifying patients who, despite being initially triaged as low-acuity, later require intensive care unit admission due to clinical deterioration. We propose an anomaly detection approach to continuously monitor patients for signs of deterioration. To achieve this, we employed a sequential anomaly detection model integrated with Long Short-Term Memory to capture temporal relationships within the data. Given the limitations of existing metrics for assessing clinical relevance, we validated the performance of our model by measuring its ability to correctly identify cases and its effectiveness in providing timely and actionable alerts to healthcare providers. Our model achieved a recall of 0.8299 and a precision of 0.8206, continuously alerting healthcare providers 36% of the time throughout the patient’s stay, with an average lead time of 3.13 hours.

In today’s travel landscape, there’s a growing demand for experiences that cater specifically to group travelers, whose needs often differ from those of solo travelers. Despite the abundance of information available on community sites like TripAdvisor, the extensive planning required can be time-consuming. This highlights the need for a recommendation system tailored to the nuances of group travel. Our study focuses on enhancing travel experiences for groups by proposing customized travel packages that take into account various preferences, such as destinations, budget constraints, and individual components like flights, hotels, and events. We introduce a method that combines Collaborative Filtering (CF) for destination recommendations with a group consensus decision-making process, factoring in individual preferences as constraints. This approach led to the creation of the GRec_Tr system, which not only suggests travel destinations but also offers comprehensive package recommendations, including flights, hotels, and activities. Our method aims to improve the overall travel experience, increase traveler satisfaction, and potentially boost sales for travel agencies. It also expands the scope of traditional CF-based systems by integrating diverse travel components into the recommendation process.

As DRAM cells continue to shrink, the conventional single error correction and double error detection (SECDED) code is not sufficient to provide DRAM error resilience. To satisfy DRAM reliability demands, various studies have proposed multi-bit error correctable ECC schemes with substantial performance and/or storage overhead compared to the SECDED code. In this paper, we propose ZEC ECC, a zero-byte eliminating compression based ECC scheme, which provides much stronger error correction capability with negligible performance overhead and no storage overhead. ZEC ECC exploits our proposed Zero-byte Eliminating Compression (ZEC) to make room for additional parity bits. Depending on the compression ratio of a memory block ( $\ge 60$ %, $\ge 50$ %, and

Computer vision has emerged as a promising tool for improving safety at construction sites through automatic scene recognition. However, traditional approaches require significant labor-intensive and time-consuming efforts for annotations. Although weakly supervised learning has the advantage of localizing objects without location information, the conventional class activation map (CAM) technique struggles with small and linear object localization and background noise at construction sites. For effective scene recognition related to construction safety, the spatial relationships between precisely localized objects are crucial. Due to the limitations of traditional CAM techniques in localizing small and linear objects, using CAM for construction safety monitoring remains inaccurate. Therefore, this study proposes a weakly supervised learning approach with an improved CAM for enhancing object detection in construction sites. The improved CAM localizes objects of various scales and is robust against background interference. Spatial relationships between localized objects are employed to determine the status of scenes for construction safety monitoring. Experiment using datasets associated with falls from ladders (FFL) demonstrates that the improved CAM surpasses the traditional CAM in mIoU (mean intersection over union) for the object localization performance as well as in the accuracy and F1-Score for recognizing unsafe scenes. This demonstrates the robust potential of employing CAM as a weakly supervised learning strategy, underlining its substantial feasibility for preventing hazards in construction sites. The proposed framework can minimize annotation efforts, demonstrating the potential of CAM as a viable computer vision technique for efficiently detecting hazards at construction sites.

Deep neural networks have shown promising results in the analysis of structural magnetic resonance imaging (sMRI) data for the diagnosis of dementia, particularly Alzheimer’s disease (AD). Different regions of the brain have diverse structures that are linked to specific functions, which could account for the variability in disease-related changes observed in sMRI scans of these areas. Understanding the overall characteristics of sMRI data is important since current popular convolutional neural networks (CNN) for deep learning do not consider the interconnection of voxels. Vision transformers have shown effectiveness in identifying long-distance connections in the brain, which has led to their success in applications such as disease detection. However, the image noise and limited scalability of self-attention mechanisms in relation to image size has hindered their widespread use in advanced Alzheimer’s analysis. To enhance information retention and reduce network complexity, this study presents a novel adaptable efficient receptive field feature extraction network. Moreover, an advanced attention mechanism with both local grid attention block and dilated global attention module has been incorporated to highlight the characteristics of AD. Next, a more improved hierarchical inverted residual feed forward network in place of multi-layer perceptron is suggested to enhance the characterization of features through the integration of information from both lower and higher layers. Finally, the global average pooling and $1\times 1$ convolution are used to reduce dimensionality, enhance non-linearity, and allow channel interactions in feature maps before being input into the classification head. The network achieved high performance in various scenarios, with average accuracies of 97.29% for AD vs. HC and 94.79% for MCI vs. HC classification using ADNI as experimental datasets.

The field of remote sensing has experienced rapid advancement owing to the widespread utilization of image sensors, drones, and satellites for data collection. However, object detection in remote sensing poses challenges owing to small objects with low resolution (LR), complex scenes, and limited data for model training. Conventional methods rely on computationally intensive models and hardware setups that are not suitable for real-time detection. To address this issue, we propose a novel sequential transfer learning method based on generative adversarial networks (GANs) that generate super-resolved data from LR for embedded systems, enabling improved performance with limited data by combining learning from both heterogeneous and homogeneous data. Additionally, we train the model sequentially, starting with the easiest data and progressing to the most complex based on the complexity levels determined by the GAN-generated images. The GAN model is trained on a diverse dataset of images and learned to generate high-resolution images from the LR, capturing finer object details for enhanced accuracy and localization capabilities. The proposed method acquires more robust features and enhances the generalizability and convergence of the model. Furthermore, the trained model of the proposed method is deployed on embedded platforms, such as Nvidia’s Jetson Nano and AGX Orin, for real-time remote-sensing object detection, with satisfactory detection performance. Evaluation metrics, such as mAP@0.5, mAP@0.5–0.95, and F1 score were used to assess the object detection accuracy. The experimental results demonstrated a significant improvement in accuracy when the proposed method was implemented with YOLOv7, achieving detection performance scores of 99.21, 98.57, 93.71, 78.38, 75.73, 48.68, 0.971, 0.971, and 0.911 on the VEDAI-VISIBLE, VEDAI-IR, and DOTA datasets, respectively.

The availability of phasor measurement unit (PMU) data enables dynamic state estimation (DSE) of synchronous generators (SGs) via nonlinear Kalman filters. The mechanical power input of an SG is not measured by PMUs. This input can be treated as three types in DSE: an unknown input, a static state variable, and a dynamic state variable that requires a first-order governor-turbine model. To properly apply DSE to power systems, the DSE accuracies for the three types of input must be compared. In this paper, an identification method for a first-order governor-turbine model is developed using frequency-domain PMU data. Then, three DSE methods corresponding to the three types of mechanical power are designed using the cubature Kalman filter. Their accuracies under PMU measurement noise and SG model parameter uncertainties are evaluated by simulating DSE of the IEEE 39-bus system. The simulation results show that all three methods provide similar estimation accuracies for the SG state but exhibit different behaviors in tracking the mechanical power. The method with the static variable is shown to have a long transient time in DSE for ambient data.

Robotic arms are increasingly being utilized in shared workspaces, which necessitates the accurate interpretation of human intentions for both efficiency and safety. Electroencephalogram (EEG) signals, commonly employed to measure brain activity, offer a direct communication channel between humans and robotic arms. However, the ambiguous and unstable characteristics of EEG signals, coupled with their widespread distribution, make it challenging to collect sufficient data and hinder the calibration performance for new signals, thereby reducing the reliability of EEG-based applications. To address these issues, this study proposes an iteratively calibratable network aimed at enhancing the reliability and efficiency of EEG-based robotic arm control systems. The proposed method integrates feature inputs with network expansion techniques. This integration allows a network trained on an extensive initial dataset to adapt effectively to new users during calibration. Additionally, our approach combines motor imagery and speech imagery datasets to increase not only its intuitiveness but also the number of command classes. The evaluation is conducted in a pseudo-online manner, with a robotic arm operating in real-time to collect data, which is then analyzed offline. The evaluation results demonstrated that the proposed method outperformed the comparison group in 10 sessions and demonstrated competitive results when the two paradigms were combined. Therefore, it was confirmed that the network can be calibrated and personalized using only the new data from new users.

The integration of unmanned aerial vehicles (UAVs) in the fifth-generation (5G) network has emerged as a promising solution to deliver seamless services and provide connectivity because of their mobile nature. In vehicular networks, connectivity between vehicles and the network can be affected by ground-level obstacles and false messages created by some deceitful vehicles. UAVs are integrated into vehicular networks to provide line-of-sight communication. This integration also poses specific challenges, such as maintaining reliable communication links and handling operational complexities related to coordination and routing, which minimizes the network throughput and packet delivery ratio (PDR) and increases energy consumption, routing overhead, and communication delay. This paper proposes improved trust analysis and artificial gorilla troops optimization (ITA-AGTO) to address the challenges associated with the security and deployment of UAVs in vehicular networks. ITA improves network security and reduces congestion by discarding incoming messages from a deceitful vehicle. It filters false messages by calculating the trust values corresponding to every message and determining whether to accept or discard the incoming message. The artificial gorilla troops optimization (AGTO) approach positions the UAVs and finds the optimal solution to transmit the data sent by the RSU or vehicle to a destination directly or through the ground base station. AGTO enhances communication intelligence and reduces the complexity of the network, thereby contributing to an overall performance improvement. The proposed approach is evaluated in terms of packet delivery ratio (PDR), communication delay, network throughput, energy efficiency, energy consumption, and routing overhead. Simulation results validate the effectiveness of the proposed approach over state-of-the-art approaches by achieving a maximum PDR (93.28%), throughput (658.17 Kbps), energy efficiency (88.19%), minimum routing overhead (865 packets), energy consumption (1.13 KJ) and communication delay (125.46 ms). The findings indicate the superior performance of the ITA-AGTO approach.

The rehabilitation of amputees has gained research attention due to digital rehabilitation systems. Traditional methods rely on intensive hospital training and prosthetics, which have limitations in accessibility and sustainability. This study proposes a mixed reality (MR) system combining real and virtual environments to enhance effectiveness and provide immediate patient feedback. The system enables amputees to perform exercises outside the hospital, ensuring continuous rehabilitation. Surface electromyography (sEMG) analysis monitors muscle activity and synchronizes intended movements with virtual prosthetics in real-time. Various algorithms filter noise from sEMG signals and reproduce amputees’ intended movements. Pre-processing includes amplification, mean filtering, and Savitzky-Golay filtering for signal quality improvement. A gated recurrent unit model with four hidden layers and dropout techniques is used for motion classification. The sEMG data was collected over 12 weeks from non-amputees and amputees, achieving 92.81% accuracy and 0.231 loss rate. Clinical trials involved quadriceps strengthening exercises, resulting in 231.61% muscle activity improvement. The MR-based system promotes muscle activation, enhances training effectiveness, and improves quality of life. This study integrates MR and sEMG technologies to provide a patient-centric, effective, and sustainable solution, overcoming limitations of existing methods. The system facilitates the convergence of technology and medical services, contributing to personalized rehabilitation frameworks.

Wheat is one of the most extensively cultivated crops worldwide that contributes significantly to global food caloric and protein production and is grown on millions of hectares yearly. However, diseases like brown rust, septoria, yellow rust, and other fungus diseases pose notable threats to wheat crops, impacting production and quality. Diagnosing these diseases is challenging, especially in areas with limited agricultural experts. Thus, creating computerized disease identification and decision-support technologies is crucial for safeguarding wheat leaf preservation and crop loss mitigation. The traditional approach to integrating data gathering and model training has substantial challenges in terms of data confidentiality, availability, and the costs related to data transmission. To address these challenges, federated learning (FL) is an appealing and effective option. Our study focuses on applying FL to classify agricultural diseases using image analysis. In our study, we conduct experiments on high-parameterized transfer learning (TL) models along with our proposed architecture based on the attention mechanism, introducing these models into a distributed learning strategy founded in FL. Our proposed architecture leverages the beneficial interactions of two cutting-edge vision transformer models including the advanced depthwise incorporating self-attention model referred to as CoAtNets, and the enhanced Swin Transformer V2, resulting in enhanced feature representation. Moreover, we introduce weight pruning into our model which is further classified by a reinforced linear attention mechanism (LA) to lower output dimensions. Our pruned lightweight (32M parameters) considerably decreases inference time with 624.249 ms and 644.899 on devices with low computational power, making it highly efficient in FL-based systems. The proposed model in our FL system significantly outperforms all other tested transfer learning models, including ConvNeXtBase, ConvNeXtLarge, EfficientNetV2L, InceptionResNetV2, ResNet152, and NASNetLarge, achieving accuracies up to 98% and 99%, precision up to 98%, recall up to 98%, and F-1 scores up to 95% across multiple input dimensions for wheat leaf disease classification.

We demonstrate that the shorter channel self-aligned top-gate (SA TG) coplanar indiumgallium- zinc oxide (IGZO) thin-film transistors (TFTs), with negative voltage applied to the back-gate, exhibit superior characteristics as driving transistors in organic light-emitting diode (OLED) pixels compared to their longer channel counterparts. The shorter channel IGZO TFTs (with a channel length (L) of 3 μm) biased with a back gate voltage of −3.5 V showed a larger subthreshold swing (SS = 0.21 V/dec) than the longer channel ones (with L = 5 μm, SS = 0.16 V/dec) with a similar threshold value (VTH = 0.7–0.8 V). A large SS is beneficial for controlling grayscale levels, especially at low gray levels, when IGZO TFTs are used as driving transistors in OLED pixels. Furthermore, the negatively back-gate-biased shorter channel SA TG coplanar IGZO TFTs exhibited significantly enhanced electrical stability compared to the longer channel ones under both positive gate bias and hot carrier stresses. The findings of this study are expected to be useful in expanding the utility of IGZO TFTs in OLED displays.

This study aimed to evaluate the effectiveness of our newly developed virtual reality head-mounted display (VR-HMD) “walker avoidance” game in reducing step-aside reaction time (SART) and enhancing agility in collision avoidance. Fifteen young adults in experimental group (EG) engaged in the “walker avoidance” game, while another 15 young adults in the control group (CG) played the “first touch” tutorial. The results showed the EG had significant decreases (p < 0.01) in both SART-standing and SART-walking when compared with pre-intervention measurements. Compared with the CG, the EG SART-standing exhibited significant decreases in both the first (p = 0.001) and second (p < 0.001) measurements post-intervention; the EG SART-walking demonstrated significant decreases in all (p < 0.05) measurements, except for pre-intervention measurement. One-dimensional statistical parametric mapping (spm1d) also demonstrated significant differences in most of the electromyography and forefoot/hindfoot ground reaction force results because the step-aside movement became quicker in the EG following training. After pushing the leg-heel contact, the EG participants made a toe-off sooner than the CG participants. Following two sessions of our newly developed “walker avoidance” game, conducted 1 week apart, the EG exhibited less collisions with virtual pedestrians and reduced reaction times to unpredictable directional change measurements compared with the CG. This study demonstrated the effectiveness of this targeted VR training program in improving motor function, which introduced a novel approach to rehabilitation as a digital therapy. It offers innovative perspectives and an approach for clinical rehabilitation, while also providing new ideas for the VR content development industry.

Rehabilitation training is essential for a successful recovery of upper extremity function after stroke. Training programs are typically conducted in hospitals or rehabilitation centers, supervised by specialized medical professionals. However, frequent visits to hospitals can be burdensome for stroke patients with limited mobility. We consider a self-administered rehabilitation system based on a mobile application in which patients can periodically upload videos of themselves performing reach-to-grasp tasks to receive recommendations for self-managed exercises or progress reports. Sensing equipment aside from cameras is typically unavailable in the home environment. A key contribution of our work is to propose a deep learning-based assessment model trained only with video data. As all patients carry out identical tasks, a fine-grained assessment of task execution is required. Our model addresses this difficulty by learning RGB and optical flow data in a complementary manner. The correlation between the RGB and optical flow data is captured by a novel module for modality fusion using cross-attention with Transformers. Experiments showed that our model achieved higher accuracy in movement assessment than existing methods for action recognition. Based on the assessment model, we developed a patient-centered, solution-based mobile application for upper extremity exercises for hemiplegia, which can recommend 57 exercises with three levels of difficulty. A prototype of our application was evaluated by potential end-users and achieved a good quality score on the Mobile Application Rating Scale (MARS).

Preventing Use-After-Free (UAF) bugs is crucial to ensure temporal memory safety. Against UAF attacks, much research has adopted a well-known approach, lock-and-key, in which unique, disposable locks and keys are first assigned respectively to objects and pointers, and then on every memory access, checked for a match. Attention has been drawn again to this approach by recent work that capitalizes on a vast abundance of virtual address (VA) space in the lock assignment, thus being able to prevent UAFs in stripped binary. However, as this VA-based lock-and-key scheme tends to rapidly consume virtual space, it is likely to suffer from high performance overhead. In this paper, we propose a new scheme, called the VA tagging, whose goal is to tackle this performance problem with the support of the Memory Tagging Architecture (MTA) introduced in several commodity processors. In our scheme, the original VA-based locks are augmented with tags of MTA. As a VA-based lock can be assigned to multiple objects with different tags, the same VA is reused for many objects without compromising temporal safety. We have observed in our experiments that this tagging scheme lowers the VA consumption rate drastically by one order of magnitude. We implement a light-weight memory allocator, Vatalloc, by modifying existing allocators, dlmalloc and jemalloc, to employ the VA tagging scheme for efficient prevention of UAFs. Our evaluation shows that Vatalloc with allocator modifications only incurs 1.70% (on dlmalloc) and 3.05% (on jemalloc) of runtime overhead without considering performance degradation of MTE. As a result of simulating the tagging architecture assuming the worst-case, postulating MTE precise trapping mode incurs performance overhead of 30.9% based on dlmalloc, and 25.5% based on jemalloc. If imprecise mode is assumed, the slowdown is measured 16.9% for dlmalloc and 12.0% for jemalloc respectively. Vatalloc only incurs 19.0% and 3.0% memory overhead for dlmalloc and jemalloc respectively.

A smart contract is a program executed on a blockchain. However, once the smart contract is deployed on the blockchain, it becomes visible to all participants and remains immutable. Thus, any sensitive information or vulnerabilities in smart contracts can be exposed to potential attackers. To protect the confidentiality of smart contracts, existing studies execute smart contracts in a trusted execution environment (TEE). However, they still suffer from vulnerabilities in smart contracts and potential memory-vulnerability problems. If an attack such as privilege escalation occurs by exploiting this vulnerability, the TEE can have a detrimental effect on the entire system as it has the most privileges in the system. To mitigate the memory vulnerability of the smart contracts in TEE, we propose MECAT, a prototype for memory-safe confidential smart contracts. In essence, MECAT runs smart contracts written in Rust, a memory-safe language, in ARM TrustZone. And MECAT is developed as a software library, allowing developers to easily apply MECAT to their smart contracts. According to our evaluation, MECAT only incurs a 1.36x performance overhead and 0.30% power overhead in single-node environments and can process the 16 clients concurrently in the network made with 8 peer nodes.

Massive jellyfish outbreaks have put human lives and marine ecosystems in great danger. As a result, the jellyfish detection methods have drawn a lot of attention, following two directions optical and sonar imaging. This work focuses on using optical imagery and CNN-based deep-learning object detection models to detect jellyfish. While labeled data of jellyfish play an important part in training deep learning models, there are a few open and available labeled datasets. Hence, we create our dataset to train these models using our model-assisted labeling method with over 11 thousand images of underwater jellyfish and corresponding annotation files in PASCAL VOC format. Our model-assisted labeling method saves the work of classical manual labeling by 70 percent, which is developed into application with YOLOv5. However, the YOLOv5 baseline suffers from the trade-off between real-time performance and low accuracy. Hence, an improved YOLOv5-nano is introduced based on adding GAM and replacing conventional Conv with CoordCov modules into the backbone of the conventional structure. The experiment results show that our improved model increases the accuracy of the conventional one by 1.3% and outperforms others including RetinaNet, SSD, Faster R-CNN, YOLOv6, and YOLOv8 at 89.1% mAP@0.5. On generalization performance, we verify the effectiveness of our work by conducting a test set of 15 different types of jellyfish with various shapes, colors, resolutions, and backgrounds. To conclude, our work establishes a comprehensive system from labeling the data, improving object detectors, and developing a feasible real-time jellyfish detector.

Non-linearities and unmodeled dynamics in the control system inevitably degrade the quality and reliability of voltage stabilization performance in DC-DC buck converters. Reinforcement Learning (RL) is an emerging method to mitigate this issue. However, traditional RL typically necessitates significant computational resources and specialized processing units, thus being an economically unreasonable option. This paper proposes a high-performance RL-based method even suitable for a cost-effective Digital Signal Processor (DSP). To address the significant challenge of time delay in a DSP when training the RL agent, this paper adopts a Real-Time Deep Reinforcement Learning (RTDRL) approach that creates an augmented virtual decision process to eliminate the delay effect. The performance is validated through software simulation (PLECS) and an actual system, through which the proposed approach demonstrated superior performance compared to existing benchmarks, including existing approaches and artificial intelligence.

This paper considers the distributed cooperative control problem for a linear stochastic multi-agent system (MAS). The optimal cooperative control design for each agent is challenging due to the limited neighboring information, contingent upon the MAS network topology. A synthesized distributed control-estimation framework is proposed to address the computationally tractable suboptimal solution. In particular, a distributed estimator extends MAS information beyond neighboring agents, allowing interactions with non-neighboring agents. The proposed control-estimation law is theoretically verified and further validated using numerical simulations.

Human action recognition systems enhance public safety by detecting abnormal behavior autonomously. RGB sensors commonly used in such systems capture personal information of subjects and, as a result, run the risk of potential privacy leakage. On the other hand, privacy-safe alternatives, such as depth or thermal sensors, exhibit poorer performance because they lack the semantic context provided by RGB sensors. Moreover, the data availability of privacy-safe alternatives is significantly lower than RGB sensors. To address these problems, we explore effective cross-modality distillation methods in this paper, aiming to distill the knowledge of context-rich large-scale pre-trained RGB-based models into privacy-safe depth-based models. Based on extensive experiments on multiple architectures and benchmark datasets, we propose an effective method for training privacy-safe depth-based action recognition models via cross-modality distillation: cross-modality mixing distillation. This approach improves both the performance and efficiency by enabling interaction between depth and RGB modalities through a linear combination of their features. By utilizing the proposed cross-modal mixing distillation approach, we achieve state-of-the-art accuracy in two depth-based action recognition benchmarks. The code and the pre-trained models will be available upon publication.

In the various industrial manufacturing processes, the automatic visual inspection system is an essential part as it reduces the chances of delivering defective products and the cost of training and hiring experts for manual inspection. In this work, we propose a new unsupervised anomaly detection method inspired by the masked language model for the automatic visual inspection system. The proposed method consists of an image tokenizer and two subnetworks, a reconstruction subnetwork, and a segmentation subnetwork. We adopt a pre-trained self-supervised vision Transformer model to use it as an image tokenizer. Our first subnetwork is trained to predict the anomaly-free patch tokens and the second subnetwork is trained to produce anomaly segmentation results from both the reconstructed and input patch tokens. During training, only the two subnetworks are optimized, and parameters of an image tokenizer are frozen. Experimental results show that the proposed method exhibits better performance than conventional methods in detecting defective products by achieving 99.05% I-AUROC on MVTecAD dataset and 94.8% I-AUROC on BTAD.

In this article, we introduce an ultra-wideband 4-bit distributed phase shifter using a lattice network. To achieve wider bandwidth, the proposed phase shifter employed an all-pass lattice network instead of the traditional low-pass ladder network. Seven cascaded 22.5° lattice phase shifters and one switched line 180° phase shifter were used to achieve 360° phase shift range. Based on our theoretical analysis, we designed the lattice network as a constant-phase shifter rather than a delay line. Implementations in the K/Ka- and E/W-bands validate the suitability of the lattice network for constant-phase shifting. Fabricated using 28-nm bulk CMOS technology, the K/Ka-band phase shifter had a size of 0.45 mm2 excluding pads. Within the frequency range of 20.5–35.5 GHz, the root-mean-square (RMS) phase error ranged from 1.6 to 5°, the RMS gain error ranged from 0.3 to 0.6 dB, and the return loss remained above 10 dB. At 28 GHz, the insertion loss was $11.6\pm 0$ .8 dB without dc power consumption. Fabricated using 28-nm FD-SOI technology, the E/W-band phase shifter had a size of 0.3 mm2 excluding pads. Within the frequency range of 63.5–100.5 GHz, the RMS phase error ranged from 2.4 to 4.6°, the RMS gain error ranged from 0.44 to 1 dB, and the return loss remained above 10 dB. At 82 GHz, the insertion loss was $11.9\pm 1$ .1 dB without dc power consumption. The proposed phase shifter demonstrated exceptional performance for multistandard operation, achieving low RMS phase and gain errors across a wide fractional bandwidth of 53.6% and 45.1%, respectively.

This paper proposes a method to reduce phantom shell effects that occur during the absorbed power density (APD) derivation using a negative permittivity material (NPM). In this study, APD is examined as a primary parameter to assess the effectiveness of the proposed method, especially following the introduction of APD above 6 GHz, in accordance with 2020 ICNIRP guidelines. As the frequency increases, the impact of the phantom shell can affect APD measurement results. To investigate this, a dipole antenna operating in the $6\sim 10$ GHz frequency range was used as the radiation source. Therefore, this study proposes a method to reduce effects by using NPM. The selection of the NPM is achieved through input impedance calculations method. Simulation results indicate that before applying the proposed method, the APD difference between the presence and absence of the phantom shell ranged from 20.6% at 6 GHz (W/ shell =100, W/o shell =82.9 W/m2) to 69.3% at 10 GHz (W/ shell=52.3, W/o shell=30.9 W/m2). After applying the method, these differences reduced to 6.9% at 6 GHz (proposed method=77.2, W/o shell=82.9 W/m2) and 13.5% at 10 GHz (proposed method=35.1, W/o shell=30.9 W/m2). This study is the first to highlight the impact of phantom shell effects on APD, revealing potential overestimations in the 6-10 GHz range and suggesting that these effects may worsen at higher frequencies. The proposed method effectively addresses these issues within current standards while providing insights for future standards development in higher frequency bands.

Accidents resulting from fires caused by electrical devices are frequent occurrences, inflicting substantial damage to both human lives and infrastructure in the Republic of Korea. To ascertain whether these fires stem from external or internal infrastructure factors, investigators such as the police, The National Institute of Scientific Investigation, and the National Fire Research Institute conduct fire-causing inspections. However, obtaining conclusive results in this regard is an intricate process, exacerbated by the dearth of adequate digital forensics tools and related programs. Among electrical devices, multi-socket outlets also contribute to fire incidents. This study explores the feasibility of employing CNN-based deep learning object detection models for fire-causing inspection systems targeting multi-socket outlets. Specifically, we introduce a novel image dataset comprising 6009 images of post-fire multi-socket outlets remaining, categorized into two classes: “burnt-in” and “burnt-out.” This dataset is utilized for training various models, including the YOLO-series (v5, v6, and v8), Faster-RCNN, RetinaNet, and SSD. Results from our experiments show the feasibility of six CNN models in detecting the cause of fire in post-fire sockets. Particularly, YOLOv5s surpasses other models with an accuracy of 89.1% mAP@0.5, a model size of 14.4MB, and an inference time of 44.5ms (equivalent to 22 fps) on RTX 3050. Subsequently, the trained models are implemented in an operational application for trial testing during an executive period.