Source Camera Identification (SCI) plays a vital role in digital image forensics by verifying the origin of images in criminal investigations and intellectual property protection. However, the growing use of digital zoom in modern imaging devices introduces geometric distortions that significantly degrade the performance of traditional PRNU-based SCI methods. Existing techniques often rely on brute-force searches over a wide range of zoom scales, resulting in high computational cost and limited robustness. In this paper, we propose a blind resynchronization method that efficiently estimates the digital zoom factor through a fast resampling-based algorithm. The estimated factor is then used to proportionally rescale the test image, enabling accurate alignment with the reference fingerprint. This approach enhances correlation reliability and reduces errors caused by scale mismatch. Experiments conducted on the ForensiCam-215K dataset demonstrate that the proposed method significantly improves identification accuracy while maintaining high efficiency, making it suitable for practical forensic applications involving digitally zoomed content.

This paper proposes a dilated semi-dense convolutional network for multichannel speech enhancement using machine learning. The baseline, SpatialNet, employs a Conformer for narrowband processing but uses a simple 3-layer CNN block, limiting local information extraction. To improve performance, we replace the Convolutional Neural Network (CNN) block with dilated convolution, dilated dense convolution, and the proposed dilated semi-dense convolution.

The special wide gap (WG) frequency-hopping (FH) sequence set with no-hit-zone (WG NHZ FHS set), combining the characteristics of both the special WGFHS set (which features stricter WG property) and the NHZ FHS set, endows the FH communication system with robust anti-interference capability, enabling superior anti-jamming performance and enhanced security. In this paper, a general method is introduced for designing WG NHZ FHS sets. By selecting different classes of sequence sets as component sequence sets, the new method can generate a new class of special WG NHZ FHS sets, the length of the NHZ for which can take even number greater than 3. The experimental results show that the new FHS sets exhibit favorable performance according to the theoretical bound.

This paper introduces a hybrid analog-digital architecture that couples a 40 nm phase-change memory (PCM) crossbar array with a lightweight FPGA-based digital processing unit through a stochastic-gradient adjustment (SGA) rule. On-chip experiments achieve 98.14% accuracy and 0.94 BPC, surpassing PCM+SGD baselines. The SGA framework thus charts a viable path toward large-scale, on-device training for edge intelligence systems.

In this note, we demonstrate that CCZ-equivalence can be applied to vectorial plateaued functions to construct new plateaued Boolean functions that are CCZ-inequivalent to every component of the original function. More specifically, by applying CCZ-equivalence to a vectorial plateaued function with exactly two amplitudes s₁ = 0 and s₂ > 0, we show that the resulting function exhibits three distinct types of components: (1) plateaued components with amplitude s₁ = 0 (i.e., bent functions); (2) plateaued components with amplitude s₂ > 0; and (3) non-plateaued components whose Walsh spectra take exactly five values. To illustrate this construction, we revisit two infinite families of vectorial functions introduced in [3]. Our results confirm the existence of numerous plateaued components in the CCZ-derived functions that are CCZ-inequivalent to any component of the original function.

In order to clarify the characteristics of input current harmonics, moreover suppress harmonics and improve power grid quality for the indirect matrix converter (IMC), the mathematical model is established based on the transfer function of IMC and the harmonic components are ascertained. Furthermore, the pulse barycentre concept, in which the amplitude, width and distribution position of a pulse are all considered, is introduced to analyze and determine the frequencies and magnitudes of main harmonic components. In addition, the principle of the output voltage vector pulse proportional symmetric distribution, which is able to attenuate input current harmonics, is discovered in the process of theoretical analysis and utilized under different modulation strategies. Simulation and experimental results show that, by using either dual space vector modulation (SVM) or two-stage pulse-width-modulation (PWM) with the pulse proportional symmetric distribution, the input currents mainly contain (6n±1)th low order harmonics, which are irrelevant to the output frequency, and the harmonic contents decrease as the frequency increases. Simulation and experimental results coincide with the theoretical results, which demonstrate the validity of the theoretical analysis.

As B5G (beyond 5G) communication techniques continue to advance rapidly, efficient and precise identification of input signals has become essential for maintaining stability in data transmission across wireless and satellite communication systems. Direction-of-arrival (DOA) estimation plays a vital role in B5G communications, especially when it comes to beamforming. In this study, we implemented the maximum likelihood (ML) algorithm for B5G DOA estimation in OpenAirInterface (OAI). In terms of hardware, we implemented it on an Intel i9-9960X CPU. During simulation testing, we estimate the performance of the ML algorithm with different numbers of antennas and compare the accuracy and time. In addition, to improve the estimation performance of the ML algorithm and reduce the time spent on DOA estimation, we further expand the single-user architecture to a multi-user parallelization architecture using multi-threading. In the multi-user parallelization architecture, we use four threads to perform four DOA estimates simultaneously. Finally, experimental results show that when the SNR is greater than -10dB, the accuracy of 16, 32, and 64 antennas can reach within 0.1 degrees. The DOA estimation accuracy is best when the number of antennas is 64 at 10dB, and the mean square error (MSE) reaches 10^-6. When the number of antennas M = 64, the average execution time of the multi-user architecture is 2485.89 ms, which is only about 1.33 times the execution time of 1875.85 ms for a single user. However, it can complete four estimation tasks simultaneously, effectively improving inference throughput.

Analog in-memory computing (AIMC) executes matrix-vector multiplications (MVMs) inside memory to alleviate the von Neumann bottleneck and improve energy efficiency. This tutorial classifies AIMC circuits in a memory-agnostic way, namely, current-domain, charge-domain, charge-redistribution, capacitive-division, resistive-division, and time-domain IMC. We explain each type of AIMC circuit with simple mathematical models. Furthermore, we review key device and circuit nonidealities (e.g., process variation, IR drop, sneak paths, and I/O quantization/ nonlinearity) with practical mitigation strategies in circuitry and peripherals. Finally, we organize hardware-aware training into three complementary families—probabilistic/precise modeling, physical modeling, and hardware-in-the-loop techniques—providing a mathematically grounded bridge between circuits and learning for robust, scalable AIMC accelerators.

Strict quasi-synchronization may be hard to maintain at all times in practical FHMA networks (e.g., infrastructureless ad hoc networks), strong no-hit-zone frequency-hopping sequence (SNHZ-FHS) could minimize the Hamming correlation for time-shifts outside of the no-hit-zone (NHZ). In this letter, theoretical bounds on the maximum partial Hamming correlation (PHC) of SNHZ-FHS are studied. The proposed bounds are strict and obtained in the literature for the first time. Moreover, Zhou-Tang-Niu-Udaya bound and Zeng-Zhou-Liu-Liu bound are the special cases of the proposed bounds.

This study proposes a method for predicting river water-level distribution using time-delay dynamic mode decomposition (tdDMD) and sparse modeling using a graph filter bank (GFB). In a previous study, sparse-coded dynamic mode decomposition on graph (SC-DMD-G) proposed by Arai et al. (2021) enabled the derivation of the time-evolution equation of water-level changes on a graph using a data-driven approach. However, SC-DMD-G has limitations in accurately representing nonlinear and complex variations in river water levels, making it challenging to improve its prediction accuracy. Therefore, tdDMD was integrated into SC-DMD-G to achieve more precise linear modeling of nonlinear dynamical systems. To validate the effectiveness of the proposed method, two types of experiments were conducted: one using artificial data and the other using real observed data. In the artificial data experiment, time-series data generated using a process-driven model, i.e., a tank model, are used to compare the reproduction performances among dynamic mode decomposition (DMD), tdDMD, SC-DMD-G, and the proposed method. In the real observed data experiment, water-level observation data collected via web scraping were used to evaluate river water-level distribution prediction during non-rainfall periods to compare the prediction performance of the proposed method with existing methods. These experimental results demonstrate that the proposed method outperforms DMD, tdDMD, and SC-DMD-G in terms of reproduction and prediction performance.

This paper proposes a generalized method for estimating the modulation period of non-cooperative signals without requiring prior knowledge of their modulation system. To achieve this, a unified mathematical model is first established that characterizes various periodically modulated signals as the convolution of a basic waveform with a periodic Dirac pulse train. Based on this model, it is theoretically derived that the signal's envelope spectrum exhibits discrete spectral lines at integer multiples of the modulation frequency, whose amplitudes are shaped by the autocorrelation of the basic waveform's spectrum. To address potential estimation ambiguities arising from unfavorable spectral line distributions, an adaptive framework is developed. This framework calculates the kurtosis of the autocorrelation of the envelope spectrum to decide whether to estimate the period directly from the envelope spectrum or via its autocorrelation waveform. Furthermore, maximum second-order cyclostationarity blind deconvolution is iteratively applied to enhance the discrete spectral lines in the envelope spectrum, thereby strengthening the periodic modulation characteristics and resolving ambiguities. Simulation results demonstrate the method's generality and effectiveness across various signal types, showing advantages in non-cooperative signal processing scenarios where the modulation scheme is unknown.

In this study, a novel image sharing technique is proposed to enhance the confidentiality and integrity of data transmission between fingerprint sensors and processors. The technique employs a specialized encryption and decryption mechanism applied to the fingerprint images, thereby ensuring secure communication over SPI or USB channels. Importantly, the proposed method preserves the performance and compatibility of fingerprint recognition, authentication, and embedded cryptosystems. An efficient and accurate encryption-decryption framework has been developed, incorporating both essential and nonessential methods. Unlike conventional approaches, which continue to exhibit security vulnerabilities in the transmission of sensor images to the CPU, the proposed scheme addresses these deficiencies directly. The effectiveness of the method has been rigorously evaluated through comprehensive experiments assessing accuracy, robustness, and reliability, including comparative analyses using good-to-gray images.

Interfacing physical world with digital world is essential in recent days to enrich our cyber physical systems. Low-power system architecture design is particularly important because of the available resource limitations, especially available power and communication bandwidth. As an efficient way to cope with limited resources the authors have proposed to “do most of the things before changing energy domains”. Local computational sensing and direct use of energy such as kinetic energy for sensing and computation have a potential to drastically reduce electrical power consumption. This article summarizes such new attempts of integrated Micro Electro Mechanical Systems (MEMS) low-power computing and sensing, with theoretical prediction towards miniaturization to High Aspect Ratio Nano Systems (and/or Structures, HARNS).

Knowledge distillation (KD) is effective for compressing object detection models into lightweight student models. However, traditional distillation losses often fail to capture the geometric structure of high-dimensional feature distributions. In this paper, we propose a Dynamic Sliced Wasserstein Approximation (DSWA) distillation framework for YOLOv8-based object detection, leveraging the Sliced Wasserstein Distance (SWD) with adaptive projection reduction. Our key innovation lies in dynamically decreasing the number of random projections during training. Starting with dense projections for accurate feature alignment, DSWA progressively sparsifies projections to suppress noise and accelerate convergence. Experiments on the BDD100K dataset demonstrate that DSWA achieves a 2.8% mAP improvement over YOLOv8. Compared to fixed-length projection SWD variants, our method reduces training time by 32%.

Recommender systems have become more crucial for informed service consumption, product selection, and decision-making in the era of overabundant information and the digitized economy. Session-based systems have emerged in recent years as a new paradigm for recommender systems. Although session-based recommenders have been extensively investigated, there are currently no unified issue statements for them nor detailed explanations of their characteristics and difficulties. In this paper, deep learning with optimal feature selection approaches are used for effective feature selection and classification. Different data related to product reviews and movie reviews are considered as input for this suggested approach. Initially, these datasets are given to the count vectorizer for converting the ”message” column's text into numbers. These converted raw data's are pre-processed utilizing similarity based data filling, min-max normalization and fuzzy c-means clustering to fill the absent values, standardization and to reduce the redundant data present in the dataset. Then, the features from pre-processed data are extracted. Aquila Optimization based approach is employed in the suggested method to decrease the number of resources required to describe a large set of data. Finally, a hybrid LSTM-SVM classifier is utilized for classification purpose. In this model, the softmax unit of the LSTM is substituted through SVM to predict the five different classes based on the customers reviews. The valuation outcomes shows that the suggested approach achieves 95%, 96%, 97% of accuracy, 90%, 91%, 93% of precision, 95%, 88%, 95% of specificity rate for three various datasets. As a result, the recommended strategy is the greatest option for a successful recommendation system.

Social networks have become essential communication channels; however, they simultaneously enable the propagation of harmful content that undermines societal well-being. Existing methods for detecting harmful posts predominantly use binary classification frameworks, which fail to distinguish between specific harmful content types and encounter significant challenges with class imbalance when extended to multiclass scenarios. In this study, we present a novel heterogeneous graph-based approach for the multiclass classification of harmful social media content, specifically addressing the multiclass imbalance problem inherent in this domain. We propose a structure-aware oversampling technique that extends the heterogeneous graph transformer architecture to identify three distinct categories of harmful content: misinformation, biased opinions, and inflammatory rhetoric. Our method generates synthetic nodes while preserving the complex interconnections characteristic of social media networks by enhancing the GraphSMOTE algorithm with network-specific constraints. These constraints maintain the semantic integrity of user-post and post-element relationships while addressing class imbalance. In extensive experiments on Japanese COVID-19 vaccine-related social media data, we demonstrated our method's effectiveness.

Fake news on social media poses serious societal risks. Therefore, early detection during the initial stages of diffusion is crucial for minimizing its impact. However, existing propagation-based methods inherently rely on sufficient diffusion, which limits their effectiveness in early scenarios. To address this challenge, we propose a novel early detection framework that integrates generalized distillation and multi-task learning. Our approach transfers knowledge from a teacher model with complete diffusion information to a student model restricted to early-stage data, while jointly optimizing fake news classification and diffusion pattern prediction as auxiliary tasks. We adopt an evolving dynamic graph convolutional network with time-dependent weights to effectively model the temporal evolution of propagation structures. In comprehensive experiments on the FakeNewsNet (PolitiFact and GossipCop) and PHEME datasets, our method achieved up to 16.4% improvement in the F1 score within one hour of diffusion onset. In ablation studies, we further validated the contribution of each component, thereby highlighting the significant impact of distillation on early-stage performance. In this study, we introduce generalized distillation in the domain of propagation-based fake news detection to offer a novel solution to the early-stage data limitation problem.

Knowledge distillation (KD) focuses on feature imitation and transfer of classification knowledge, but it is inefficient in extracting localization information. To address these issues, this paper proposes curriculum learning distillation (CLD) to transfer classification knowledge from KD to CLD. Furthermore, we apply CLD to the anchor-free detector FCOS (CLD-FCOS), improving object detection accuracy by introducing center-ness based on localization distillation. Experimental results on the MS COCO dataset demonstrate the versatility of CLD method, which can consistently improve the distillation performance of various detectors.

Conventional chip quality assessment predominantly depends on manual inspection, a process that is not only time-consuming and labor-intensive but also susceptible to human error. This study investigates the utilization of ResNet50, a deep convolutional neural network (CNN), in semiconductor chip quality assessment. By training the ResNet50 model on a large-scale dataset of chip images encompassing both defective and non-defective samples, we seek to develop a highly precise automated system for defect classification. Our methodology comprises preprocessing chip images, employing data augmentation techniques to improve model generalization, and fine-tuning ResNet50 through transfer learning. Experimental results indicate that Our method achieves an average accuracy of 98.83%, which means the ResNet50-based model surpasses traditional inspection methods in both speed and reliability. The proposed approach holds substantial implications for semiconductor manufacturing, providing a scalable, reliable, and efficient solution for quality control. Future research could expand this model to real time defect detection systems embedded within production lines, enhancing overall manufacturing efficiency and min-imizing costs.

Vertical federated split learning (VFSL) is a technique in which multiple clients and a single server cooperate to train a deep neural network model. In VFSL, the DNN model is split between the clients and the server. They exchange intermediate features and gradients of the loss at the split boundary to perform both training and inference. This approach enables effective model training while preserving data confidentiality, as it avoids directly sharing input data and labels directly. In this paper, we investigate an adversarial examples (AEs) generation attack in a VFSL setting, where clients and the server continue to collaborate during the inference phase. Specifically, a malicious attacker, who is one of the clients, manipulates the intermediate feature sent to the server so that it behaves as AEs to mislead the inference results. To generate AEs, the attacker needs to obtain the gradient of the loss (calculated from the inference result and the ground-truth label) with respect to the intermediate feature. However, during inference, the server does not transmit gradients. Accordingly, the attacker trains an attack model using the intermediate features and gradients available during the training phase, which is then used to estimate the sign of the gradient from the target intermediate feature during inference. A variant of the fast gradient sign method (FGSM) algorithm is used to generate AEs. Our experimental results demonstrate that the generated AEs significantly degrade inference accuracy compared to perturbations generated by random noise.

Card-based protocols are cryptographic protocols that use a deck of physical cards. In this paper, we deal with card-based protocols using a standard deck of playing cards, which are commonly and commercially available. For finite-runtime committed-format protocols, Mizuki (CANS 2016) proposed an eight-card AND protocol with four random bisection cuts and a six-card COPY protocol with one random bisection cut. In this paper, we propose a partial-open action, which reveals any position of the face of cards, by generalizing the half-open action introduced by Miyahara and Mizuki (IJTCS-FAW 2022). Using the partial-open action, we propose four-card AND protocols with three random cuts, a four-card COPY protocol with three random cuts, and a four-card base conversion protocol with a random cut. We note that, without partial-open actions, these AND and COPY protocols with random cuts only are known to be impossible to construct using four cards. Therefore, the partial-open actions are inherently necessary to obtain our results.

Infrasound monitoring is regarded as promising for the early detection of disasters because infrasound, which is often generated by strong natural events, propagates across long distances through the atmosphere at the speed of sound. An important obstacle is that no available sensor device can measure the entire frequency range of infrasound at low cost, which hinders the construction of a sensor network for infrasound monitoring. Fusion sensing might be one workaround to alleviate this hurdle to future development. This paper presents a proposal for the fusion of a MEMS pressure sensor and a miniature infrasound microphone, with subsequent investigation of optimal signal processing to merge the outputs of these two sensors having different characteristics. The proposed method is based on a gradient descent algorithm combined with deep unfolding. After the hyperparameters of the gradient descent algorithm are optimized using deep learning under suitable constraints, the filter coefficients are derived through parallel observation with a high-precision reference sensor using an arbitrary signal containing a wide frequency band. These study findings demonstrated that the proposed fusion sensing achieves approximately 26% reduction in RMSE compared with a simple combination of low-pass and high-pass filters or a direct sum of the two sensor outputs.

As an alternative to two-dimensional codes, there is growing interest in techniques that embed information in images and detect that information from their printed counterparts. In this paper, we propose the detection of information from data-embedded images on curved surfaces without a priori knowledge. In our proposed method, auxiliary lines are added in advance around the image in which data are embedded and information is detected by correcting the image using these lines during capture. This method mitigates limitations related to image-capturing conditions in conventional approaches.

This paper describes a deep-learning-based methodology for assessing arteriosclerosis through the automated detection of carotid artery calcification in dental panoramic radiographs. The proposed framework introduces a detector architecture that sequentially integrates a segmentation module and a classification network, achieving both computational efficiency and high diagnostic accuracy. A TransFuse-based segmentation model is combined with a ResNet classification backbone, enabling a substantial reduction in the number of parameters compared with conventional approaches while maintaining high detection precision.

Images taken outdoors at night are captured under various conditions such as light sources, and it is difficult to improve the visibility of high-contrast images, especially in low-light environments. In this study, we propose an algorithm that combines bright pixels (e.g., light sources) and dark pixels (e.g., background) using multiple methods and fuzzy sets to make high-contrast images in low-light environments easier to see by enhancement of the dark areas while decreasing the intensity of the bright areas. Experimental results show that the proposed method is effective by comparing it with conventional methods.

There are two types of image compression methods which are lossy compression and lossless compression. However, lossy compression has a problem of missing information despite its high compression efficiency. In contrast, lossless compression has a compression efficiency of about 40% at maximum, on there hand is used for data distribution because it can restore the original information. Lossless compression methods include dictionary methods that calculate the frequency of occurrence of the entire data and methods that use data conversion and encoding. However, all of these methods have the problem that the conversion process, encoding, and dictionary creation are time-consuming. In addition, when implemented in hardware, the logic circuits become complex. This paper proposes a new lossless compression method that optimizes the code length according to the data appearance interval. We have confirmed that the proposed method has a compression ratio equivalent to that of conventional lossless compression methods, and that it has a simple hardware configuration.

Recent advancements in Convolutional Neural Networks (CNNs) have led to significant increases in computational complexity and power consumption. Approximate computing, especially through the use of approximate multipliers, has emerged as a promising approach to realize the reduction of power, area and delay by sacrificing some computational precision. Approximate multipliers produce results with errors, thus their effects to CNN accuracies need to be evaluated. However, the lack of efficient software tools for simulating approximate multipliers hinders their widespread adoption. This paper proposes ApproxTorch 2.0, a high perfomance simulation framework based on PyTorch and CUDA for evaluating 8-bit signed integer based approximate multipliers for CNNs. To realize higher performance, CUDA and Nvidia GPUs are used to accelerate the simulation process. Two approximate layers are implemented: 2D Convolution layer and Linear layer. The simulation of approximate multiplications is done by accessing pre-defined look-up tables (LUTs) stored in GPU memory. By dedicated optimization in the customized CUDA approximate GEMM kernel, ApproxTorch 2.0 achieves 171× speedup compared with CPU simulation and 17.2× speedup compared with previous ApproxTorch 1.0 on running ResNet50. Lastly, gradient estimation is added to support the retraining of approximate CNNs with quantized values and proved to be effective for recovering the accuracy loss after approximation.

A distributed network system is a dynamical system in which multiple subsystems are connected through a physical/ communication network. There are many applications such as power systems. In this paper, we propose a method of event-triggered model predictive control for distributed network systems with switching topologies. In the proposed method, each subsystem is controlled by state-feedback using its state and the state of neighbors. Its gain is calculated by the upper controller (supervisor), only when a certain event-triggering condition is satisfied. The design problem of state-feedback gains is reduced to an LMI (linear matrix inequality) optimization problem. Finally, the proposed method is demonstrated by two numerical examples.

Accurately predicting energy consumption of the heating ventilating and air conditioning system is crucial for achieving building energy conservation. To overcome the limitations of traditional methods in terms of generalization capacity, this study proposed a novel prediction method for the HVAC system energy consumption based on the ensemble learning model. In this method, the operation variables of chilled and cooling water systems (i.e., supply/return temperature and flowrate) and environmental variables (i.e., temperature and humidity) are utilized as inputs, three single models (i.e., support vector regression, extreme learning machine, and decision tree) are employed as the base models, and the Gaussian process regression is utilized as the stacked model. Experiments are conducted on the HVAC system of the public building, and the performance of the ensemble learning model is verified using practical measured data. Results show that the ensemble learning model is superior to the single models in predicting cooling capacity, heat dissipation and total power. More importantly, the ensemble learning model can provide reliable confidence intervals, effectively quantifying the uncertainty of the prediction results.

As the global energy mix shifts toward renewable energy, microgrids, as a key enabler for efficient energy utilization and the integration of distributed generation, face complex and ever-changing energy management challenges. The uncertainty of the output of distributed power sources (such as solar and wind power), the interactions between devices, and the multi-objective optimization requirements within microgrids make it difficult for traditional energy management methods to achieve accurate forecasting and efficient scheduling. To address these challenges, this paper proposes the TGD-RL model, an innovative approach that combines deep learning techniques. This model integrates three advanced techniques: the Transformer, the Graph Neural Network (GNN), and the Deep Q-Network (DQN). The Transformer module utilizes a multi-head attention mechanism to capture long-term temporal dependencies in microgrid data, making it suitable for processing data with time-series characteristics. The GNN uses graph convolution operations and node embedding techniques to model the topology and dynamic interactions of each device in the microgrid. The DQN uses a state-action-reward mechanism to continuously optimize energy management strategies and achieve efficient scheduling decisions. Experimental results on two public datasets, PJM and MISO electricity market price data and NREL wind and solar data, demonstrate that the TGD-RL model outperforms other baseline models in energy forecast accuracy, achieving mean absolute percentage errors (MAPEs) of 6.5% and 5.8%, respectively, representing reductions of 36.3% to 39.0% compared to the optimal baseline model. The operating costs of the microgrids were reduced to 1,250 yuan and 1,180 yuan, respectively, representing decreases of 17.4% to 37.8%, while energy self-sufficiency increased to 78% and 82%, respectively, representing increases of 10.0% to 26.0%. Ablation experiments further validated the essential role of each component in the model's performance. This research demonstrates that the TGD-RL model can effectively address complex energy management issues in microgrids, providing a new technical path for improving the economic efficiency, stability, and energy self-sufficiency of microgrids. It also holds significant implications for the development of smart grids and the efficient use of renewable energy.

This paper proposes a distributed dual decomposition algorithm for solving mixed-integer linear programming (MILP) problems in multi-agent systems. In the proposed approach, a MILP problem is transformed into an approximated problem by linear programming relaxation, which enables each agent to independently solve their local subproblems. The theoretical analysis provides guarantees on both the feasibility error bounds and the optimality gap of the solutions obtained by the proposed method. The effectiveness of the proposed method is shown through a numerical example of a fairness-aware multiple traveling salesman problem.

In this paper, we propose applying bit-serial arithmetic units to reduce the circuit area of neural network inference engines. Additionally, we propose applying datapath pipelining and zero skipping to significantly reduce the required clock cycles. In recent years, studies have demonstrated the efficacy of neural networks in voice and image recognition applications; however, an extremely large number of multiply-and-accumulate operations are required in order to achieve high accuracy. Therefore, we explored the application of bit-serial arithmetic units to these operations to reduce circuit area. Bit-serial arithmetic is a method of sequentially calculating multi-bit data by inputting and outputting one bit at a time, which enables the reduction of the circuit area and amount of wiring. The disadvantage of this method is that it requires a large number of clock cycles. For example, a bit-serial multiplier with an input of N bits requires 2N cycles. In this study, pipeline processing and zero skipping were applied to reduce the required clock cycles. Zero skipping reduces the required clock cycles by skipping the calculation of an input activation when the value of that activation is zero. We propose two methods of zero skipping: reactive zero skipping, which checks whether activation is zero before the bit-serial operation starts, and proactive zero skipping, which reads ahead, examining subsequent memory locations, during the bit-serial operation and skips all consecutive zeros in one step. The effectiveness of zero skipping is highly dependent on the ratio of zeros in the input activation. In a convolutional neural network (CNN) that uses a rectified linear unit (ReLU) as the activation function, the input activation of the second and subsequent convolution layers has a high ratio of zeros. To further increase sparsity and improve the effectiveness of zero skipping, we propose setting the dropout rate during training as high as possible without affecting the recognition accuracy. We implemented a CNN using the proposed bit-serial arithmetic units and a CNN using conventional parallel arithmetic units, and compared their performances. The former exhibited a 22.9% smaller circuit area than the latter. In addition, the increase in the number of required clock cycles was limited to 2.12 times, and the clock period was reduced by 47.4%, resulting in a 7.8% reduction in runtime.

Modulation is crucial for multiplexing, reducing bandwidth, and improving transmission efficiency. By incorporating neural networks, we can optimize modulation for specific communication channels. We propose two neural network optimized modulation schemes: a regular constellation mapping for 16-ary modulation and an irregular mapping for 2^m-ary modulation. These maintain gradient flow during backpropagation, allowing adjustments to constellation points to minimize bit error rates (BER) while keeping system complexity manageable. The results show our polar-coded modulation schemes outperform traditional uniform QAM with about a 0.5 dB gain under low SNR. Additionally, these schemes can also be applied to LDPC-coded modulation systems to improve BER performance.

Specific emitter identification plays a crucial role in the field of information security. To improve identification performance in complex electromagnetic environments, this letter proposes a dual-branch network, FADC-Transformer, which combines frequency-aware dynamic convolution (FADC) and Transformer. It can adaptively fuse frequency-domain features from FADC and time-domain features from the Transformer. Specifically, FADC introduces a frequency-band attention mechanism and dynamic kernel generation, which can dynamically adjust convolutional kernel parameters according to the inputs, resulting in better robustness. Experimental results show that the accuracy of FADC is improved by 16% compared with static convolution, and the dual-branch structure significantly enhances identification performance.

Private information retrieval (PIR) is a mechanism for retrieval of a message while keeping its index secret. Tian's and Sun's methods are representative PIR methods, which are mainly evaluated in terms of download rate. In this paper, we propose a new method based on Shamir's secret sharing. Furthermore, a comprehensive evaluation including communication cost, computation time, and message size is performed; our results demonstrate that the proposed method outperforms conventional ones.

In this article, the authors propose an electronic, linear, continuously tunable phase shifter (PS) with an adjustable phase range of leading +180° to delay -180° within the phase shift bandwidth (PSBW) and a maximum adjustable phase range of +240° to -240° within the half power bandwidth (HPBW). This PS enables flexible design of its operating frequency and PSBW and maintaining a constant linear slope of -116.6° per susceptance (jb) for any design PSBW. The PS consists of two different wavelengths of transmission lines and four identical LC parallel resonators (LC tanks), with phase variation achieved by adjusting the capacitance values, susceptance values, or resonance frequencies within the LC tanks. The authors also present an analysis of the electrical theory underlying this PS. Theoretically, the transmission coefficient T ranges from 0dB to -0.973dB, and the reflection coefficient Γ peaks at -6.99dB. Additionally, the authors provide design guidelines for the PS. Based on these guidelines and available lumped components, a PS operating at 3GHz with a PSBW of 400MHz was developed. Finally, measurement results confirm a generally consistent between theoretical predictions and practical implementation. This demonstrates that the proposed PS can be designed with arbitrary center frequencies and PSBW, exhibiting excellent S-parameters and linear continuously tunable phase shifting performance.

Although a sampling rate and a resolution of a time-interleaved A/D converter (ADC) have improved remarkably by recent advance of digital calibration, it is still difficult to achieve an effective resolution larger than 12 bits for sampling rate higher than 1 GS/s. This limitation is mainly caused by higher-order effect of sampling-timing mismatch among unit converters. To overcome the fundamental limitation, fully digital calibration of a time-interleaved ADC with cascaded higher-order sampling-timing correction is presented. In the proposed correction method, by using a reference ADC as an only additional analog component, analog tuning is eliminated, allowing mismatch effects to be corrected solely through post-digital processing. Due to its fully digital nature and unlimited correction in principle provided by the cascaded processing, accuracy is only limited by digital implementation cost, which is mitigated significantly with CMOS scaling. The extension to a sub-sampling time-interleaved ADC is also presented for a broad range of applications. Effectiveness of the proposed calibration was verified by extensive simulation with the 3rd-order sampling-timing correction for both standard and sub-sampling time-interleaved ADCs as well as measurement of a prototype time-interleaved ADC, which proved 11.5-bit effective resolution (71.2-dB SNDR) at 1GS/s with the 2nd-order correction.

This study examines data-driven attack detectors that utilize maximum likelihood estimation for cyber-physical systems. The proposed methodology optimally estimates the input-output trajectories of these systems using both pre-experimental data and real-time measurements, in accordance with the principles of maximum likelihood. The attack detector identifies the presence of an attack by comparing the estimated trajectories with actual measured trajectories. The theoretical contributions of this study include the demonstration of a fundamental limitation, specifically, the set of undetectable attacks when disturbances are negligible. Furthermore, when disturbances can not be disregarded, the proposed method can detect attacks with a specified false rate based on the detectable condition derived through a χ² test.

Many researchers expect to apply microrobots in narrow environments for tasks such as exploration and maintenance. However, digital control, which is the primary robot control method, faces computational cost and circuit miniaturization challenges. The authors have been studying neuromorphic circuits, which mimic biological neural functions for robot control, acting as a central pattern generator (CPG) as a driving circuit to perform the walking motion. Previously, we constructed a neuromorphic circuit on an integrated circuit and we successfully implemented the neuromorphic integrated circuit in millimeter-scale microrobots. However, the microrobot lacked sensory input, which prevented the robot from adapting to the robot's movement in response to external environmental changes. This paper proposes a neuromorphic integrated circuit capable of adaptively switching the gait of an insect-type microrobot in response to light stimuli. The proposed circuit incorporates a receptor cell model that mimics biological sensory neurons, enabling the transformation of external light input into electrical signals using photovoltaic cells (PV cells). The electrical signals are processed through synaptic and CPG models to switch locomotion patterns. The authors systematically measured gait patterns to evaluate the operating range while varying the power supply voltage of the receptor cell model and the output voltage of PV cells. We observed the results, which clarified the regions where stable wave gait, tripod gait, or unstable outputs. Furthermore, we measured the I-V characteristics of a PV cell. Also, we confirmed that its output voltage matches the designed switching threshold of the proposed circuit, enabling optical control without additional signal conditioning. These findings demonstrate that the proposed circuit can be a low-power, sensor-responsive gait controller for future autonomous microrobots.

We propose a method for reliability research at the circuit-level chips. The degradation model used combines the Wiener process and the Arrhenius acceleration model. The degradation data analyzed is the pin leakage current sampled during the constant stress acceleration degradation test of Flash memory chips at different temperatures. This method has low testing costs while providing a comprehensive reflection of the degradation conditions of the tested samples. In this work, we established the model through mathematical derivation, and then estimated the distribution of the model parameters by generating bootstrap samples. Then, under the premise of completing the model accuracy test, we completed the estimation of the remaining life of the sample through Monte Carlo simulation.

Integrated sensing and communication (ISAC) enables simultaneous wireless communication and environmental sensing, where joint beamforming design is key to balancing both tasks. To address the complexity and trade-off challenges of joint beamforming design in multi-user multi-target ISAC systems, this letter proposes the cross-attentive Pareto transformer (CAPT), an end-to-end deep learning framework that integrates enhanced spatial embeddings and cross-attention to jointly optimize beamforming. By leveraging Pareto multi-task learning, CAPT efficiently generates the Pareto front of solutions in a single inference. Simulation results show that CAPT achieves better Pareto front quality and generalization than weighted minimum mean square error (WMMSE) and convolutional neural network (CNN)-based baselines.

Compression by Substring Enumeration (CSE), which is one of the lossless data compression algorithms, and various versions of CSE have been proposed. In encoding of CSE, substrings of given fixed length and their frequencies within circular string for an input string are output as a codeword. The circular string is made by connecting the first symbol and the last symbol of an input string. In decoding of CSE, the circular string is reconstructed from its substrings and their frequencies. Furthermore, the minimum length of substrings for which the decoding does reconstruct the circular string has been proved, together with a reconstruction algorithm. However, the algorithm requires substrings to have no errors.

Therefore, in this paper, we propose an error correcting algorithm which can detect one of the substrings having one bit-flipping, one bit-insertion, or one bit-deletion error and correct the bit error. By applying the proposed algorithm, we can reconstruct a circular string from a set of substrings and their frequencies including only one substring which has at most one bit error.

Cayley hash functions, which are hash functions based on walks on Cayley graphs of groups, are a well-studied class of hash functions with provable security (under assumptions on computational hardness of some group-theoretical problems). In a recent work by Aikawa, Jo, and Satake (Transactions on Mathematical Cryptology, 2023), they proposed a variant of Cayley hash functions called left-right Cayley hash functions, whose design intended the problem of finding a collision to be as difficult as the problem of finding collisions of two Cayley hash functions simultaneously. In this paper we show, as opposed to the expectation, that finding a collision of their hash function is reduced to finding a collision of a single Cayley hash function. We also propose a possible countermeasure against this issue.

With the development of artificial intelligence technology, the demand for multi-domain pose estimation in complex scenes is growing. The existing single model solution faces the problem of insufficient ability in cross domains of multi-scene, which is caused by the different numbers of keypoints and complex features of samples. To solve this problem, we propose a new method that has a four-stage pose estimation framework. This framework applies methods such as object detection, domain classification and pose estimation. As the experimental results show, on the testing set of mixed domains, the accuracy of our method is 5.1% higher than the best one of the existing methods, which ensures high performance pose estimation in many applications.

This paper introduces a latent port-Hamiltonian framework using deep learning to improve the robustness for vision-based control. Although reinforcement learning and deep learning are promising solutions to control system states with differentiable policies, physics-free methods usually suffer from unstable and low-confident results with respect to the system dynamics. We propose a vision-based control architecture by employing a port-Hamiltonian model in the latent space of autoencoder (AE) to achieve physically consistent control. Specifically, we apply a variational autoencoder (VAE) to encode visual observations into a low-dimensional latent space, where the port-Hamiltonian energy structure is learned. Moreover, we introduce AI-Pontryagin, which generates control signals similar to optimal control inputs through a neural network inspired by optimal control theory. The experimental results show that our method achieves more accurate and stable control performance compared to baseline approaches.

In order to terminate iterations earlier of the linear programming (LP) decoder based on the alternating direction method of multipliers (ADMM) (ADMM-LP), this letter proposes an early termination (ET) criterion based on the number of parity-check constraints of iterative solution vector. The criterion can be used to detect erroneous codewords in advance and thereby avoid unnecessary iterations, without increasing the additional computational complexity. Compared with existing termination criteria of ADMM-LP decoding algorithm, the proposed ET criterion can significantly reduce the average number of iterations at low signal-to-noise ratio (SNR) regions with little effect on the decoding performance.

In stabilizer-based quantum secret sharing schemes, it is known that some shares can be distributed to participants before a secret is given to the dealer. This distribution is known as advance sharing. It is already known that a set of shares is advance shareable only if it is a forbidden set. However, it was not known whether any forbidden set is advance shareable. We provide an example of a set of shares such that it is a forbidden set but is not advance shareable in the previous scheme. Furthermore, we propose an advance sharing scheme for stabilizer-based quantum secret sharing of quantum secrets such that any forbidden set is advance shareable.

Classic McEliece has gathered attention as a candidate in NIST post-quantum cryptography standardization. However, it suffers from high demands on the decryption algorithm, making it unsuitable for resource-constrained devices. In this paper, we propose a novel implementation method, named giant footprint sharing, that reduces memory size during decryption in Classic McEliece. The decryption algorithm processes a large number of intermediate variables computed from a secret key in memory. The giant footprint sharing identifies the largest variable among them and allocates a memory-sharing structure to store it, thereby reducing the overall memory size regardless of the implementation platform. The giant footprint sharing can also be combined with existing acceleration techniques, such as fast Fourier transformation. We evaluate Classic McEliece with the giant footprint sharing on the Arm Cortex-M33 CPU and show that it reduces memory size by up to 46% without significant degradation in computation time compared with the existing fast-implementation by Chen et al. (at TCHES 2021). Extensive experiments with the giant footprint sharing further reveal that it maintains a constant memory size regardless of the compiler optimization, and it also achieves an optimal balance in the trade-off between memory size and computation time. The giant footprint sharing is remarkable for any scheme, that contains a large-scale matrix computation and the life cycle for each variable is limited.

A Periodic-Finite-Type shift (PFT) is a set of bi-infinite sequences that prohibit the appearance of forbidden words in a periodic manner. More precisely, a PFT $\cX_{\{T, \tilde \F\}}$ is a set of bi-infinite sequences $\x$ characterized by a period $T\in \N$ and a family $\tilde\F=(\tilde\cF^{(0)},\tilde\cF^{(1)}, \cdots, \tilde\cF^{(T-1)})$ of indexed finite sets of forbidden words $\tilde\cF^{(0)},\tilde\cF^{(1)}, \cdots, \tilde\cF^{(T-1)}$, so that the $r$-shifted sequence $\sigma^r(\x)$ of $\x$ does not contain words in $\tilde\cF^{(i \mod T)}$ at position $i\in \Z$. The study on PFTs is strongly related to the study on constrained systems with unconstrained positions, which have the property as both error-correcting codes and constrained codes.

The capacity of a PFT is an important value that gives us the maximum coding rate when a random sequence is encoded to a sequence in the PFT. In this paper, we derive the capacity of a PFT in two ways, using the fact that an arbitrary family $\tilde \F$ is transformed into a family $\F=(\cF^{(0)},\emptyset \cdots, \emptyset)$, where each forbidden word in $\cF^{(0)}$ has the same length $k$, so that $\cX_{\{T, \tilde \F\}}=\cX_{\{T, \F\}}$. When $k \le T$, the first proof derives the capacity directly from the definition, and the other proof does from block partitioning of the adjacency matrix of a certain graph representing $\cX_{\{T, \F\}}$. We also present a partial result on the capacity when $k>T$.

Accurately predicting demand and adapting to rapid market changes are common difficulties for enterprise supply chain management. Traditional forecasting methods and even many existing deep learning models often fail to capture complex dependencies across multiple temporal features. This results in limited accuracy and delayed adjustments, which directly impact inventory, logistics, and profitability. To overcome these issues, we introduce a puzzled BiLSTM (PZ-BiLSTM) model, a specialized deep learning architecture designed for supply chain forecasting. Instead of traditional BiLSTM, our approach integrates structured feature blocks and temporal alignment techniques, which allows the model to identify both short-term fluctuations and long-term trends more effectively. The simulation of the model is performed using the publicly available Walmart sales dataset. When compared with existing demand forecasting models, our puzzled BiLSTM achieved an R² of 0.9708, demonstrating its superior predictive performance. This model not only improves forecast precision but also enables real-time adjustment in supply chain decisions by combining the strengths of bidirectional sequence modelling with a dynamic puzzle-based feature integration strategy.