Side-channel attacks (SCAs) have been reported to reveal secret keys by analyzing the power consumption or electromagnetic leakage on cryptographic circuits. Deep-learning SCAs (DL-SCAs) have been actively studied due to their superior performance over conventional attack methods. Conventional attacks use statistical models between power consumption and the internal register values (e.g. the Hamming distance (HD) over consecutive clock is exploited). The attacker can guess the secret key from the internal register values estimated from power consumption. Similarly, neural network (NN) models are constructed to predict the internal register values from the power consumption in the case of a DL-SCA. However, it has been reported that NN models cannot be trained well by using HD as labels because the frequency distribution on each HD is greatly biased. We propose a method of mitigating the class imbalance problem using a conditional variational encoder. We report the results of a DL-SCA using this method against hardware-implemented AES. The method was able to reveal all the partial keys in less than 400 waveforms. Attack performance in this result was better than previous work.
A major issue in the conventional First-Order Reduced and Controlled Error (FORCE) learning architecture is the low processing speed due to extensive matrix-vector calculations for learning. In this study, we present the field programmable gate array architecture of FORCE learning that achieves the processing of 1-Msps 500-Node data. As a result, the learning was accomplished approximately 3,400 times faster while maintaining the original accuracy in a simulation under specified conditions.
In this study, the effects of different packet reception timings on the packet success rate (PSR) are investigated in a 3-mode-based MC-CDMA on-demand WSN system with a packet composition method. In particular, PSR in the carrier model, which takes into account the code phase difference between pseudonoise codes and the phase difference between carrier waves, is evaluated. Results indicate that the PSR performance is significantly degraded compared with the conventional evaluation results with baseband transmission, in which only the code phase difference between pseudonoise codes is considered. It is also found that there are two steps that improve PSR: increase the number of slots to be combined (Ncomb) and increase the number of nodes to be woken up (Nwk).
In Japanese next-generation digital terrestrial television broadcasting(DTTB) for ultrahigh-definition television(UHDTV) broadcasting, the application of the layered division multiplexing-band segmented transmission-orthogonal frequency division multiplexing(LDM-BST-OFDM) transmission scheme is studied. Sequential decoding and direct decoding are proposed as decoding schemes. In addition, gray mapped LDM(GLDM) is also proposed as a mapping scheme. Reception Performance of sequential decoding and direct decoding have not yet been compared. GLDM and normal LDM have also not yet been compared. In this paper, Sequential decoding and direct decoding against the lower layer(LL) required carrier-to-noise ratio(CNR) of the LL are discussed. As seen from a computer simulation, the reception performance of direct decoding is better than that of sequential decoding in some cases, where the required CNR of LL is smaller than that of upper layer(UL) in direct decoding. In addition, GLDM can improve reception performance compared with normal LDM when direct decoding is used in LDM.
Bit-depth expansion is a technique for recovering a high-bit-depth (HBD) image from a low-bit-depth (LBD) image. Although recent high-performance displays can display each color in 10-bit or 12-bit depth to express more detailed color gradations, most existing image and video sources still have an 8-bit depth. Simple restoration from an LBD image into an HBD image leads to artifacts such as missing high-frequency information and false contours, so high-quality bit-depth expansion approaches have been explored. In recent years, deep convolutional neural network (CNN) techniques have provided excellent performance in many image processing tasks such as super-resolution, classification, deblurring, and denoising. In this paper, we propose a novel bit-depth expansion approach using predictive filter flow (PFF). PFF predicts a spatially variable filter using the CNN and reconstructs the target image by filtering it with the input image. The proposed method trains the PFF network that predicts the filter to transform an LBD image into an HBD image. Experimental results confirm that the proposed method provides better results than state-of-the-art bit-depth expansion algorithms.
In the classification of sign language motion, it is vital to measure the movements of many points on human body. While the method of calculating the nodal coordinates of human body from camera image is limited to the detection of two dimensions, acceleration sensor can detect three-dimensional movements. In this investigation, on the measurement of sign language motion, we used a backscatter communication system, which enables synchronized multi-channel reception of sensor signals from battery-free sensors. Acceleration sensors were attached to four locations on both wrists and elbows, and data were acquired from four signers. SVM was used as the classifier, and the classification accuracy was evaluated for 20 sign language words by cross-validation. The experiment demonstrated that classification accuracy of 86.0% to 91.0% could be obtained using this setup.
We describe a leaf region detection and leaf area calculation algorithm for estimating the growing conditions of small tomatoes. In leaf region detection, a color image taken by a depth camera is analyzed by image processing. The leaf region is divided into more than one hundred quadrangle meshs, and the spatial coordinates of the vertices of the mesh are calculated from the depth information obtained by the depth camera. To reduce the effect of measurement noise, an approximate surface expressed by a two-variable quadratic function is obtained, and the leaf area is calculated as the sum of the area of the meshs on the approximate surface. In the experiment, the algorithm will be evaluated using leaf measurements of small tomatoes in a greenhouse.
In this paper, we propose a cloud distribution prediction model in which fully convolutional networks are used to improve the prediction accuracy for photovoltaic power generation systems. The model learns the cloud distribution from meteorological satellite images and predicts the cloud image 60 min later. We examined the applicability of Day Microphysics RGB as input to the cloud image prediction model. Day Microphysics RGB is a type of RGB composite image based on the observation image of Himawari-8. It is used for daytime cloud analysis and can perform detailed cloud analysis, for example, the discrimination of cloud areas such as upper and lower clouds. The performance of the proposed method is evaluated on the basis of the root mean square error of the prediction and ground truth images.