Search

The ultimate goal in adversarial defense is to build a universally robust defense against all types of attacks, but ongoing arms race between adversarial attacks and defenses show the difficulty in building a deterministic defense to work towards the goal. Leveraging the idea of a mixture model, in this paper, we introduce a new Randomized Adversarial Defense method (RAD) to increase the robustness against adversarial examples. RAD is designed as a simple, yet effective random mixture of a global model and one or more local models. The mixture is able to create a random decision boundary for a test instance, making it harder for an adversarial example to succeed, and thus increasing the robustness of the defense. The global model is adversarially trained to provide the baseline robustness. The local models are aimed to supplement the global model and thus the decision boundary of a local model is expected to be adjacent to the decision boundary of the global model. These models then together form a random mixture to create a randomized (non-deterministic) decision boundary for each test instance at the end. Such a randomization scheme reduces the adversarial risk since the adversary has to approximate the best attack despite of the given complete knowledge of the parameters of the individual classifiers. By proposing the notion of having a global and local models with different focuses in the mixture and the way of creating a local model which has minimum dependency on the base (global) model, RAD provides a simpler and more flexible, yet effective approach building a randomized defense, compared with the existing randomization based methods. Experimental results show that our simple randomization approach outperforms the most robust deterministic defense method and performs competitively upon the existing randomized defense method against strong adaptive attacks on CIFAR10 and CIFAR100. Refereed/Peer-reviewed

Microsatellite instability (MSI) is a clinically important characteristic of colorectal cancer. Standard diagnosis of MSI is performed via genetic analyses, however these tests are not always included in routine care. Histopathology whole-slide images (WSIs) are the gold-standard for colorectal cancer diagnosis and are routinely collected. This study develops a model to predict MSI directly from WSIs. Making use of both weakly- and self-supervised deep learning techniques, the proposed model shows improved performance over conventional deep learning models. Additionally, the proposed framework allows for visual interpretation of model decisions. These results are validated in internal and external testing datasets.

Cortical connectivity analysis is a widely used method for understanding the causes of neurological disorders and related brain mechanisms. Although there exist numerous activity analysis toolboxes for functional near-infrared spectroscopy (fNIRS), there are only a few cortical connectivity analysis toolboxes. In 2019, we released a MATALB toolbox named OptoNet , which has helped researchers to analyze brain networks using fNIRS. In this study, we developed an advanced MATLAB toolbox, named OptoNet II, to add new features that overcome the shortcomings of OptoNet . With these new features, OptoNet II can efficiently analyze cortical connectivity according to brain region using any fNIRS channel sets and can present the results of two connectivity analyses with auto-thresholding based on surrogate tests. To evaluate the efficacy of the new functions, the finger-tapping task experiment was carried out before and after transcranial direct current stimulation (tDCS) in the primary motor area. OptoNet II can efficiently show the effects of tDCS on functional brain region connectivity, which has been difficult to confirm by conventional methods. In this article, we propose the OptoNet II as a useful and efficient toolbox for researchers who want to perform cortical connectivity analysis using fNIRS.

To detect the three-phase current in the complex plane of a DC link shunt inverter, an algorithm for restoring the current is required. In this paper, a method of dividing the detection voltage and the compensation voltage to match the output voltage as much as possible to reduce the total harmonic distortion while restoring the current is proposed. In addition, an overmodulation algorithm for a 12-step output, which corresponds to the largest voltage in a DC link shunt inverter, is proposed, and a current recovery method in the overmodulation region is proposed. To determine how to ensure a linear output voltage, the fundamental frequency of the output voltage is analyzed through a Fourier series, and a new voltage vector whose fundamental frequency amplitude is equal to the amplitude of the command voltage is calculated. Finally, the performance of the proposed algorithm is verified through simulation and experimentation. The output of a motor was increased by using overmodulation, and the harmonics of the current based on the output voltage were analyzed through a Fourier series.

Federated learning (FL) is an emerging technique that trains machine learning models across multiple de-centralized systems. It enables local devices to collaboratively learn a model by aggregating locally computed updates via a server. Privacy is a core aspect of FL, and recent works in this area are advancing the privacy guarantee of an FL network. To ensure rigorous privacy guarantee for FL, prior works have focused on methods to securely aggregate local updates and provide differential privacy (DP). In this paper, we investigate a new privacy risk for FL. Specifically, FL may frequently encounter unexpected user dropouts because it is implemented over a large-scale network. We first observe that user dropouts of an FL network may lead to failure in achieving the desired level of privacy protection, i.e., over-consumption of the privacy budget. Subsequently, we develop a DP mechanism robust to user dropouts by dynamically calibrating noise with account of the dropout rate. We evaluate the proposed technique to train convolutional neural network models on MNIST and FEMNIST datasets over a simulated FL network. Our results show that our approach significantly improves privacy guarantee for user dropouts compared to existing DP algorithms on FL networks.

LiDAR sensors have the advantage of being able to generate high-resolution imaging quickly during both day and night; however, their performance is severely limited in adverse weather conditions such as snow, rain, and dense fog. Consequently, many researchers are actively working to overcome these limitations by applying sensor fusion with radar and optical cameras to LiDAR. While studies on the denoising of point clouds acquired by LiDAR in adverse weather have been conducted recently, the results are still insufficient for application to autonomous vehicles because of speed and accuracy performance limitations. Therefore, we propose a new intensity-based filter that differs from the existing distance-based filter, which limits the speed. The proposed method showed overwhelming performance advantages in terms of both speed and accuracy by removing only snow particles while leaving important environmental features. The intensity criteria for snow removal were derived based on an analysis of the properties of laser light and snow particles.

Herein, we unveil the deposition and etch mechanism of GeCl 4 on the SiGe surface. At the high temperature, GeCl 4 is dissociated to GeCl 2 and then worked as a deposition source. Thus, the rate determinant step of surface growth is GeCl 4 dissociation, and a novel precursor that quickly dissociates to GeCl2 will be a proper precursor target to Ge growth at the low-temperature process. Otherwise, at the low temperature, GeCl 4 works as an etching gas by reacts with surface Ge/Si atoms and forms Ge$_{2} H_{n} {Cl}_{6-n}$ or GeSiH$_{n} Cl_{6-n}$ (n=2,3) molecules. From the etch mechanism analysis, the first activation energy of Ge desorption is lower, 65.8 (kcal/mol), than GeCl 4 dissociation (101 kcal/mol), but the etched surface has higher energy, -6.7 (kcal/mol), than the Ge doped, -19.2 (kcal/mol). This energy profile successfully explains the experimental observation, deposition at high temperature, etch at low temperatures. Additionally, we figured out that the GeCl 3 intermediate shows the most tightly bind to surface atoms.

Physically unclonable functions are the most secure technology to keep a static cryptographic key in a device due to its unclonable and volatile properties, where PUF responses disappear when the power is off. These make a PUF-based key is intrinsically secure against reverse engineering. Though the PUF itself is secure, its integration into a system must be carefully designed to keep its security. This paper presents SAMPUF™ design approach and its integration with security systems.

Implementing binary neural networks (BNNs) on computing-in-memory (CIM) hardware has several attractive features such as small memory requirement and minimal overhead in peripheral circuits such as analog-to-digital converters (ADCs). On the other hand, one of the downsides of using BNNs is that it degrades the classification accuracy. Recently, ResNet-style BNNs are gaining popularity with higher accuracy than conventional BNNs. The accuracy improvement comes from the high-resolution skip connection which binary ResNets use to compensate the information loss caused by binarization. However, the high-resolution skip connection forces the CIM hardware to use high-bit ADCs again so that area and energy overhead becomes larger. In this paper, we demonstrate that binary ResNets can be also mapped on CIM with low-bit ADCs via aggressive partial sum quantization and input-splitting combined with retraining. As a result, the key advantages of BNN CIM such as small area and energy consumption can be preserved with higher accuracy.

An object's shape may be deformed by a combination of gravitational, hydrostatic, mechanical, and electromagnetic forces. Therefore, to predict the deformation, it is necessary to know each force's distribution inside the object. Various expressions and methods, such as the Lorentz, Kelvin, generalized, and Korteweg-Helmholtz forces, can be used to calculate the electromagnetic force on a dielectric or magnetic material. However, the distributions of the aforementioned forces inside materials may differ significantly. We adopt the concepts of infinitesimal particles and external electromagnetic fields to address this issue. Adopting these concepts enables the electromagnetic force densities inside dielectric or magnetic materials to be uniquely determined. We refer to this type of density as the locally defined electromagnetic force density (F LEM ). This study primarily focuses on the derivation of F (LEM )• Subsequently, the distribution of F LEM is then demonstrated using simple numerical models.

Predicting stock prices are difficult as they are affected by diverse and complex factors. Therefore, among the various indicators that affect stock prices, key indicators must be selected. Hence, we present a new feature selection method using a multivariate convolutional neural network model to select key indicators that affect stock prices. In addition, we used data of daily net buying and net selling amounts based on investor type, unlike technical indicators or financial data used in other studies. The proposed feature selection method is validated by comparing the predicted accuracy of the stock price using selected and overall indicators. Furthermore, we compare the data and verify the sector using the more efficient data by analyzing industrial sectors.

This paper introduces a new approach to forecast daily stock trends by merging structured and unstructured datasets. This study intends to reveal the effectiveness of using supplemental datasets for accurate prediction of stock prices. A set of features, which is seemingly highly correlated with daily stock price variations, are selected using random forest optimization technique. Stock-relevant keywords that are extracted from news articles are converted into a time-series dataset in terms of temporal frequency. Convolution neural network (CNN) based deep learning models are generated separately for stock trading data and keyword frequencies from news articles, and two CNN models are merged together for training input datasets. The analysis results show that merging two different datasets may generate the better forecasting results than using stock trading datasets only. Additional issues for future analysis and implementations are discussed.

Images taken outdoors are highly likely to generate noise due to rain, snow, and fog. So, removing noise is one of the important fields in image processing. This field usually requires a real-time, high-speed image processing. The noise removal field could be used in the pre-processing stage of extensive image processing such as perception processing of autonomous vehicles. Therefore, optimization for real-time processing should be preceded and an approach that characteristics of color space will be one of them. This research applies several color spaces to image processing and analyzes them through three steps. First, the dataset construction. A ground truth and a noised dataset are needed for quantitative evaluation. Second, image de-noise. Third, evaluation through indicators. Through this, analyze the possibility of optimizing image processing through color space.

This paper introduces and controls a new variable radius pulley that passively changes the effective radius of the drive pulley in a belt driven transmission in response to changing load torque. The mechanism used in this paper is based on the principle that the effective radius of the pulley increases passively when torque is applied to the pulley by the cam element and the elastic element. A belt-driven transmission with a conventional pulley of constant radius has a fixed maximum torque and speed, but the variable radius pulley in this paper changes the maximum speed and torque to enable more efficient driving. This paper explains the principle of operation of the new variable radius pulley and introduces the experiments that performed the position control of the robot joint by applying the variable radius pulley.

While the performance of convolutional neural networks (CNNs)-based single image super-resolution (SISR) has been greatly improved, the enormous parameter sizes and computational complexity of the underlying CNNs make hardware implementation difficult. Recently, several lightweight SISR methods have been developed, but they still do not consider various structural problems that may occur in hardware implementation. To solve this problem, we propose a slice-based SR using light-weight network (LWN) and a slice-based SR using LWN with relation loss (LWNRL). First, LWN(RL) adopts a slice-based architecture to facilitate system-on-chip (SoC) implementation. Second, LWN(RL) avoids global connection modules that are not suitable for SoC implementation, with minimal performance penalty. Finally, we propose a new loss to improve the performance of LWN without additional cost. Experimental results show that LWNRL achieves significant efficiency of SR model. Especially, the larger the resolution or scale factor, the better the performance of LWNRL than the conventional methods.

We propose multi-modal anomaly detection in embedded systems using time-correlated measurements of power consumption and memory accesses. Time series of power consumption of the processor and memory accesses between L2 cache and memory bus under known-good conditions are used to train one-class support vector machine (SVM) and isolation forest classifiers. These side channels have complementary anomaly detection capabilities. Experiments on a high-fidelity processor emulator show that the method accurately detects anomalies.

In order to apply the model predictive control to the power converter, a study on the calculation amount reduction algorithm was performed in various topologies such as 2-level, 3-level, multi-level converter. Similarly, in order to apply the model predictive control to the 3-level 4-leg converter, it is necessary to study the calculation amount reduction algorithm. In this paper, instead of considering 81 candidate voltage vectors for every sampling period as in the conventional model predictive control, the optimal switching state is selected considering only 7 candidate voltage vectors located near the reference voltage vector. The sector, prism, and tetrahedron are selected sequentially by using the position of the reference voltage vector, and the preselected 7 candidate voltage vectors are the vectors constituting the tetrahedron. The proposed method represents an improved model predictive control which reduces the amount of computation and does not affect performance. This method constructs the 3-level 4-leg NPC inverter simulation and experimental setup to compare the performance with the conventional method.

Protection of information is of paramount importance in today's digital age. Physically Unclonable Functions (PUFs) are considered a secure method for security key generation because they generate responses that exist only during operation. A challenge regarding the use of PUFs is to achieve high reliability. Therefore, various schemes such as temporal majority voting [2], [3], [4], spatial majority voting [1], BCH [1], [3], and burn-in [3], are applied to improve the stability of the responses. While a recent paper proposed a method of oxide-break to achieve zero error [5], it is controversial if it is a real PUF since the response value (i.e. the status of the oxide-break) can be observed by reverse engineering. Automotive is an application area where reliability is particularly important, as failures may lead to critical accidents. To satisfy the reliability of AEC-Q100 Grade 1, functionality under −40-to-125°C in ambient temperature (Ta) must be guaranteed, even considering the aging effects on a chip. To satisfy IS026262 ASIL-B, the fault coverage must be over 90%. This paper shows a PUF satisfying both AEC-Q100 Grade 1 and IS026262 ASIL-B, where our testing temperatures cover −40-to-150°C in junction temperature (Tj) to compensate for the increased thermal heat within the SoC package.

There is enormous demand for high-bandwidth DRAM: in application such as HPC, graphics, high-end server and artificial intelligence. HBM DRAM was developed [1] using the advances in package technology: TSV, microbump and silicon-interposer. Owing to these advances, HBM has a much higher bandwidth, at a lower pin speed rate, than conventional DRAM. However, the 3D-stack structure causes TSV interface and PDN problems: TSV connection failure and 3D-accumulation of IR drop, which increases the total cost of HBM. Moreover, as memory bandwidth increases DRAM architectural challenges arise, power consumption and associated thermal problems increase as well.

In this paper, an efficient GNSS (Global Navigation Satellite System) integration to NB-IoT chipset and its solution for constrained tracking scenario (i.e. time sharing between NB-IoT and GNSS) are explained. Considering the infrequent data transfer to network in the target application of NB-IoT, it makes sense to constrain tracking scenario in order to optimize system architecture in NB-IoT SoC (system on chip) for low-cost GNSS feature support. For efficient SoC design architecture, NB-IoT and GNSS may share main hardware resources such like CPU, memory, and system peripherals. However, the time sharing of CPU for such hardware environment requires additional software solution to manage context save/restore, firmware binary switch, and data transfer between NB-IoT and GNSS. In NB-IoT, long period of eDRX (extended DRX) and PSM (Power Save Mode) modes allow enough spare time for GNSS to run exclusively on the shared resources without missing any paging request during location search operation by GNSS.

This paper introduces a new approach to forecast daily stock trends using the random forest technique. This study intends to include as many features as possible to hopefully describe various aspects of stock market trends. A number of features are selected for forecasting the trends of stock prices. The new algorithm adjusts optimal learning parameters during the data training process. The usefulness of the proposed algorithm is demonstrated by processing two stock datasets while analyzing its forecasting accuracy. Additional several technical issues for future implementations and analysis are suggested.

Data storage is one of the hottest discussion topics in today’s connected world. The amount of data growth is expected to be exponential, while budget and space remain constricted. Since the transformation of storage device from planar NAND to 3D V-NAND [1], the areal density of semiconductor storage devices has continuously evolved and has surpassed the density of magnetic hard drives. By providing the largest storage capacity in the smallest footprint, 3D V-NAND has been leading the data center revolution in recent years. However, 3D-technology scaling faces several technical challenges [2]. (1) As the number of WL stacks increases the channel-hole etch process becomes a limit, since the total WL-mold height increases. (2) Interference between cells increases since the distance between WLs becomes smaller. (3) Faster data transfer speeds are required to support higher IO bandwidth.

This paper presents comprehensive analysis result on technical patents about automated interpretation of glaucoma. Valid and relevant patents are carefully selected from the WIPSON DB, and comprehensive patent trends are analyzed. The contribution of AI techniques for automated glaucoma interpretation is mainly discussed. Several patent issues for generating future patents are suggested.

Based on the lessons learned from the accident at Fukushima NPP, many countries prepared and carried out the mitigation strategies to cope with ELAP accidents. This paper analyzed one of the coping strategy for ELAP, which is the operation of mobile DG and its pump loads. In case of mobile DG operation, the voltage and frequency of Class 1E 4.16kV safety bus could be oscillated by motor starting current, when the motor pumps connect to Class-1E 4.16kV safety bus in sequence. Thus, this paper reviewed the voltage and frequency dynamic responses of Class 1E 4.16kV safety bus and verify the adequacy of the mobile DG operation and motor pumps starting, according to the assumed ELAP coping scenarios. This paper found that the large capacity motor load starts later, it could cause the vulnerable conditions, such as motor staring failure, motor loads trip, motor mechanical stress, etc.

Intrinsic and extrinsic defects play a major role in determining solar cell efficiencies of lead halide hybrid perovskite absorbers. Here, we present the results of first-principles computations performed on MAPbBr 3 (MA= methylammonium) to study the energetics and defect levels of intrinsic point defects, namely vacancy, self-interstitial and anti -site, and extrinsic Pbsubstitution defects. While vacancies are the lowest formation energy intrinsic defects and create shallow transition levels, a number of extrinsic defects can have comparable formation energies under desirable chemical potential conditions. Therefore, carrier concentrations may be tunable with these substituents. Further, some extrinsic defects create deeper transition levels which can potentially be exploited to enhance solar cell efficiencies via sub-gap absorption.

In this paper, we used a PCB substrate that used a white silk legend printed on a heater pattern in the process of detecting fluorescence with an RT-PCR chip fabricated on a PCB substrate. However, since the fluorescence brightness deviations between chips were large, accurate quantitative analysis was difficult. Therefore, we fabricated chips using 4 kinds of PCBs that can be coated on the PCB substrate and experimented to select the optimal chip for fluorescence detection and obtained the results.

An important technical challenge associated with using OLEDs is to develop an appropriate statistical modeling and reliability assessment for precise lifetime prediction. We, herein, propose a modified stretched exponential decay (MSED) model for luminescence decay with respect to intrinsic emissive layer, which is dependent initial luminescence behaviors and subsequent degradation in the constant current stress tests. By using the model well fitted to experimental data from accelerated stress tests, we successfully demonstrate that a MSED extracted from statistical modeling enables precise lifetime prediction with respect to process variation and duty factor in real operation conditions.

This paper studies computing the l ∞ -induced norm of multi-input/multi-output (MIMO) linear time-invariant (LTI) discrete-time systems, which is called the l 1 analysis problem. Because such a computation involves an infinite-dimensional matrix, this paper employs a truncation idea to approximately deal with the l 1 analysis problem. More precisely, this paper develops two methods for computing the l ∞ -induced norm, in which upper bounds and lower bounds of the l ∞ -induced norm are derived on the truncated treatment. Furthermore, it is shown that the gaps between the computed upper and lower bounds converge to 0 at an exponential rate of N, where N is the truncation parameter. Finally, a numerical exampled is provided to demonstrate the effectiveness of the developed computation methods.

This paper presents an application of the command shaping method to industrial servo systems. The command shaping method, such as input shapers and bi-quad notch filters, is a well-known control technique for reducing residual vibrations. An input shaper cancels out the vibration using the convolution of a sequence of impulses with a reference command. A bi-quad notch filter places the filter zeros near the resonant poles of the closed-loop system and places highly damped filter poles. This paper compares the performance of three command shaping methods, which are zero vibration (ZV) input shaper, ZV and zero derivative (ZVD) input shaper, and a bi-quad notch filter. Numerical simulations are performed in various frequency ranges to measure the performance of the command shaping methods. Experiments are also performed using an industrial servo system to show that the command shaping methods successfully remove the residual vibration of the system, and have similar results for the various ratio of estimated frequency to actual frequency as in the numerical simulations.

This paper presents an automated technique to detect optical disc (OD) regions and to locate clipping circles to separate OD and non-OD regions using fundus images. After surveys on different OD detection techniques, a set of image and geometric processing techniques is selected and implemented. Several public fundus images with different ophthalmologic diseases are used to experiment and to verify the performance of the proposed algorithm. The proposed algorithm tends to locate clipping circles properly by enclosing OD regions with fundus images of healthy patients. However, the algorithm is not good enough to process fundus images of different ophthalmologic conditions. The overall performance of the proposed algorithm is discussed along with several experimental results. Several future research issues are also addressed.

This paper presents a method to enhance robustness in sliding mode control (SMC) combined with decoupled disturbance compensator (DDC). More specifically, a gain selection method for the DDC in industrial servo systems with unknown inertia is developed. The SMC with DDC method is a well-known robust control method which is effective for slowly-varying and bounded disturbances. However, when the gain of the DDC is improperly selected, a large error of the inertia parameter can lead to poor performance such as a large overshoot in the motor position. This paper shows that the disturbance caused by an uncertain inertia can produce a pair of complex conjugate eigenvalues of the error dynamics in the z-domain, which results in the overshoot in both position and velocity. In addition, utilizing this analysis, a gain selection method of DDC is proposed when the ratio of the actual inertia to the nominal inertia is assumed to be bounded. The analytical results are further supported using experiments performed on industrial servo systems with a ball-screw load and a belt load.

Detecting obstacles on the road is crucial for the advanced driver assistance systems. Obstacle detection on the road can be greatly facilitated if we have a vertical road profile. Therefore, in this paper, we present a novel method that can estimate an accurate vertical road profile of the scene from the stereo images. Unlike conventional stereo-based road profile estimation methods that heavily rely on a parametric model of the road surface, our method can obtain a road profile for an arbitrary complicated road. To this end, an energy function that includes the stereo matching fidelity and spatio-temporal smoothness of the road profile is presented, and thus the road profile is extracted by maximizing the energy function via dynamic programming. The experimental results demonstrate the effectiveness of the proposed method.

Manufacturing demands have increased in recent years, requiring a tack time a few tens of milliseconds to assembly machines with vibrational modes of less than a couple of hundreds of Hz. For typical high-frequency range (approximately 800 Hz) vibrations, notch filters with a depth of 1.0 are successfully used for vibration suppression. However, low-frequency range (approximately 200 Hz) vibrations typically result from the viscoelasticity of couplings and cannot be compensated for using a typical notch filter with a depth of 1.0. In addition, the damping levels can change significantly depending on specific application types as well as load and environment conditions, which in turn makes pre-implementation, off-line tests impractical. Therefore, it is desired to identify in real-time the damping conditions and adjust the notch filter depth, as well as the center frequency of the notch filter. Many methods are available for estimating the center frequency and the damping ratio in real-time, but the well-known zero-vibration derivative (ZVD) method to estimate the damping ratio cannot be applied in real-time. This paper develops a new real-time method to adaptively change the notch filter depth, using the methods of random decrement and peak detection. Experiments on an actual industrial servo system are used to demonstrate that damped vibrations are successfully compensated for by the proposed method.

This paper presents an application of a discrete derivative method with a new frequency mapping technique for adaptive-notch-filter (ANF) based resonance suppression in an actual industrial servo system. An ANF includes a frequency estimator which estimates the frequency of input signal in real time. A new discrete derivative method, which utilizes a modified bilinear transform, was recently developed to improve the frequency estimation performance of ANF and applied to computer simulation cases. The frequency estimation performance in steady state is improved, especially in high frequency ranges by using this derivative method. However the previous method can cause large estimation errors when estimated frequency varies rapidly. This paper develops a new frequency mapping technique which compensates for the frequency warping of the modified bilinear transform, and applies it with the discrete derivative method to implement an ANF. This method shows improved frequency estimation performance even when the estimated frequency varies rapidly. Numerical simulations are performed for the cases which an adaptation gain, the amplitude of an input signal, or the frequency of an input signal is high, respectively. Note that the rapid change of estimated frequency occurs in these cases. Experiments using an industrial servo system are also performed to show the improved frequency estimation performance of the developed ANF.