Technical Program

14:15–14:30 An Algorithm of Set-Based Differential Evolution for Multidimensional Knapsack ProblemAWARD Yuto Kuninaga (Fukuoka Institute of Technology), Michiharu Maeda (Fukuoka Institute of Technology)

Abstract

This paper presents set-based diﬀerential evolution (SDE) for multidimensional knapsack problem (MKP). As MKP is an NP-hard combinatorial optimization problem, many metaheuristic approaches have been addressed large scale instances. However, they frequently struggle with search stagnation and premature convergence to local optima. To deal with these issues, we present a new mutation strategy that enhances exploration capabilities within the search space. We conduct numerical experiments on the MKP instances in OR-Library. Experimental results demonstrate that our algorithm shows to be eﬀective and outperforms conventional algorithms on MKP benchmark instances.

14:30–14:45 Modeling Heterogeneous Asymmetry in Nominal Square TablesAWARD Ayaka Yoshida (Tokyo University of Science), Kouji Tahata (Tokyo University of Science)

Abstract

We propose the Clustered Conditional Difference Asymmetry (C-CDAS) model for analyzing heterogeneous asymmetry structures in nominal square contingency tables. The model extends the conventional Conditional Difference Asymmetry (CDAS) model by partitioning off-diagonal cell pairs into clusters and assigning a cluster-specific asymmetry parameter to each cluster. Model selection is formulated as a partition optimization problem over the space of all possible partitions. Since exhaustive search becomes computationally infeasible because of the Bell-number growth of candidate partitions, we develop an Agglomerative Hierarchical Clustering (AHC) algorithm based on information criteria. Applications demonstrate the effectiveness and practicality of the proposed approach.

14:45–15:00 A Time-Division Method for Transportation Planning Problems Based on Weight DesignAWARD Konosuke Ogura (Yamaguchi University), Ryosuke Adachi (Yamaguchi University), Yuji Wakasa (Yamaguchi University)

Abstract

This study introduces a route optimization method for a transportation planning problem involving a single agent, which utilizes a transition system and time-series modeling. Traditional optimization approaches require the prediction horizon to encompass the entire planning horizon, resulting in significant computational demands. To mitigate this challenge, we develop weights in the objective function that facilitate effective route selection, even with a reduced prediction horizon. The proposed time-division method decreases the computational load while preserving route quality.

15:00–15:15 Improvement of Time-Series Memory Capability Using Multilayered HiPPOAWARD Keigo Nagatsu (Yamaguchi University), Ryosuke Adachi (Yamaguchi University), Yuji Wakasa (Yamaguchi University)

Abstract

This study proposes a method for enhancing time-series memory capabilities through the application of a multilayered The high-order polynomial projection operator (HiPPO) framework. HiPPO compresses time-series data into finite-dimensional coefficient vectors by employing function approximation with orthogonal bases. Although HiPPO demonstrates superior memory capability and enhances learning performance when incorporated into Recurrent Neural Networks (RNNs), it tends to smooth out high-frequency components and local variations during the compression of long-term time-series data, leading to a loss of temporal information. To mitigate this issue, this study proposes a Multilayered HiPPO, which seeks to improve performance by maintaining temporal information through a multilayered structure.

15:15–15:30 Speeding up the CUR Factorization Algorithm Considering Singular Value DecayAWARD Haruto Ishizaki (Fukuoka Institute of Technology), Takafumi Miyata (Fukuoka Institute of Technology)

Abstract

CUR factorization is useful for obtaining a low-rank approximation of a large-scale data matrix. This small approximation has a part of columns and rows of the original matrix and preserves the important structural properties of the matrix such as sparsity and nonnegativity. The columns and rows are selected from the original matrix using singular value decomposition (SVD). The number of selected columns and rows is dynamically adjusted by exploiting the decay pattern of singular values, and the computation of SVD is required iteratively. To speed up the CUR factorization algorithm, we consider computing a partial SVD, which is only necessary part for the selections. Numerical experiments show that our approach can reduce the computational time of the algorithm while keeping the accuracy of the matrix approximation.

[EA]

Engineering Acoustics (1)

Fri, Oct 30 14:15–15:15Rooms 5,6 (4F)

Chair: Yuma Kinoshita (Tokai University)

14:15–14:30 Improving the Accuracy of a Stereo Microphone 2D DOA Estimation System Using an Additive Forgetting Coefficient Initialized According to Device RotationAWARD Kazuki Yamada (Shibaura Institute of Technology), Kenji Muto (Shibaura Institute of Technology)

Abstract

Abstract― In this study, we developed a two-dimensional (2D) directional visualization system using a stereo microphone attached to a smartphone. This system enables 2D direction estimation using only a stereo microphone by rotating the device. The purpose of this study is to improve the accuracy of direction estimation in a 2D directional visualization system. We propose an additive operation with a forgetting coefficient that resets the cumulative value in response to rotational motion. The experimental results showed that the proposed method reduced the standard deviation (SD) by 2.4°.

14:30–14:45 Effect of Peak Frequency Mismatch on Frequency-Scaling SSNMF for Separating Robust Cicada Noise from Traffic NoiseAWARD Satoshi Takei (Shibaura Institute of Technology), Kenji Muto (Shibaura Institute of Technology)

Abstract

This study investigates how differences in peak frequency (Δf) between training and target cicada noise affect source separation performance in traffic noise measurements. We compare a conventional Semi-Supervised Nonnegative Matrix Factorization with amplitude-domain additive basis deformation (SSNMF-ABD) and a frequency-scaling SSNMF method. Experimental results suggest that the proposed method was less sensitive to Δf, particularly for the spectral peak corresponding to the syllable /n/, and showed less variation in performance than the conventional method. These results suggest that the proposed method may provide more stable separation performance against individual variations in cicada noise.

14:45–15:00 Individual estimation of Head-related transfer function with anthropometric features using deep operator networkAWARD Mikiharu Ikeda (Tokyo Denki University), Ryosuke Onizawa (Tokyo Denki University), Gen Sato (Tokyo Denki University), Izumi Tsunokuni (Tokyo Denki University), Yusuke Ikeda (Tokyo Denki University)

Abstract

The head-related transfer function (HRTF) represents the transfer function from a sound source to the eardrums and contains essential information for accurate sound localization. Therefore, it has been widely applied in virtual reality and metaverse technologies. Because HRTF strongly depends on individual anthropometric features, individualized HRTF must be measured. However, measuring HRTF is difficult because it requires specialized equipment and measurement environments. Consequently, recent studies have focused on estimating individualized HRTF. In this study, we propose an estimation method for individualized HRTF with anthropometric features using Deep Operator Networks (DeepONets). Simulation results demonstrate that the proposed method outperforms conventional approaches.

15:00–15:15 Active Noise Control Based on 2D Sound Field Interpolation Using a Fast Physics-Informed Learning ModelAWARD Aoi Tokushige (Tokyo Denki University), Hayato Komaba (Tokyo Denki University), Izumi Tsunokuni (Tokyo Denki University), Yusuke Ikeda (Tokyo Denki University)

Abstract

Conventional multichannel active noise control requires error microphones to be placed within the control region, which restricts the listener's movement. To address this issue, this paper proposes an ANC method based on sound field interpolation with the Physics-Informed Extreme Learning Machine, a fast learning model that incorporates physical laws. The proposed method estimates sound pressure at unobserved locations from signals measured by a small number of microphones and uses the estimated sound pressure as the desired signal for updating the control filter. Simulation results demonstrate that the proposed method achieves noise reduction performance comparable to that of the conventional FxLMS algorithm.

[ITS]

Intelligent Transport Systems

Fri, Oct 30 14:15–15:15Room 7 (4F)

Chair: Kazuyuki Miyakita (Niigata University)

14:15–14:30 A Method of Gridlock Prevention for Traffic Signals at Intersections Using Constrained Multi-Agent Reinforcement LearningAWARD Masaru Unno (Soka University), Daisuke Kasamatsu (Soka University)

Abstract

While AI-based traffic signal control, particularly Reinforcement Learning (RL), is widely studied, real-world implementation is hindered by safety concerns. This study addresses intersection vehicle stops, a major cause of gridlock. This paper proposes a Constrained Deep Reinforcement Learning (CDRL) approach to guarantee intersection clearance while maintaining traffic efficiency. Our simulation study showed that the proposed method reduces in-intersection stops by 97% compared to fixed-time signals and by 89% compared to unconstrained baselines.

14:30–14:45 Scalable Infrastructure Assistance for Occluded Permissive Left Turns via Logistic-Regression GatingAWARD Fuma Sawa (The University of Osaka), Hiroki Nishikawa (The University of Osaka), Ittetsu Taniguchi (The University of Osaka), Takao Onoye (The University of Osaka)

Abstract

Occluded permissive left turns benefit from infrastructure sensing, but full intervention validates every left-turn vehicle against opposing through traffic at each control step. We introduce a logistic-regression gate that predicts whether exact validation is worth invoking. The gate never issues GO by itself; any release still requires the validator, otherwise the vehicle keeps its local fallback. In held-out SUMO scenarios, the gate yielded zero collisions and a +0.14 s mean left-turn travel-time difference relative to full intervention, while reducing cumulative infrastructure execution time by at least 96.6% and exact-validator calls by at least 97.0%.

14:45–15:00 Cross-Weather LOWO Evaluation of an Edge-Density Parking Occupancy Detector: Per-Slot Failure AnalysisAWARD Aaron William Chimkono (Hokkaido University), Hiroshi Tsutsui (Hokkaido University), Takeo Ohgane (Hokkaido University)

Abstract

An edge-density parking occupancy detector is evaluated under Leave-One-Weather-Out on a balanced 4-weather subset of the Snowy Parking Dataset (44 images per weather condition, 3 seeds). Drizzle-fold F1 is 77.50 % versus 95.10 % for a supervised MobileNet baseline, a 17.6 pp gap driven entirely by precision (64.63 % vs. 96.27 %). Per-slot analysis attributes 96 % of drizzle false positives to five spatially-fixed slot regions of interest from two recurring sources: background features (snow accumulations, wet/dry pavement transitions) and adjacent vehicle spillover. The largest-connected-component fraction discriminates these spurious detections from genuine vehicles with Cohen’s d=1.16, identifying spatial-coherence filtering as the natural mitigation direction.

15:00–15:15 A Consideration on Security Issues of Information Floating Kazuyuki Miyakita (Niigata University), Ryuji Koyama (Niigata University), Keisuke Nakano (Niigata University)

Abstract

Information floating (IF) is a kind of epidemic communication, where information is disseminated by direct communication between mobile nodes and their movement. In IF, nodes send information only within a transmittable area (TA). Because the performance of IF deeply depends on the position and size of the TA, TA specification is very important. Therefore, if TA information is erroneous or falsified, the performance of IF can degrade significantly. This paper considers the scenario in which an attacker who falsifies TA information. We evaluate the effect of TA falsification on IF performance using computer simulations. We also propose some methods to address the falsification and show their effectiveness.

[SIS]

Smart Info-media System (1)

Fri, Oct 30 14:15–15:30Rooms 201-203 (2F)

Chair: Nico Surantha (Tokyo City University)

14:15–14:30 Chrominance Correction for Dichromats in YPQ Color SpaceAWARD Ayaka Fujita (Yamaguchi University), Noriaki Suetake (Yamaguchi University)

Abstract

Dichromats have difficulty in the discrimination of specific color combinations. A purpose of this study is to develop a color conversion method that improves color contrast for dichromats based on change of chrominance in simulation of dichromatic vision. A proposed method improves color discrimination in dichromatic vision by adding color information which is lost in dichromatic vision to other color component in YPQ color space. To verify the effectiveness of the proposed method, it is compared with conventional color conversion methods for dichromats both qualitatively and quantitatively in experiments.

14:30–14:45 Backlit Image Enhancement Based on Image Fusion Using LDA, Cuboid and Mean FiltersAWARD Miyu Tsubone (Yamaguchi University), Ayaka Fujita (Yamaguchi University), Takanori Koga (Kindai University), Noriaki Suetake (Yamaguchi University)

Abstract

Generally, backlit images which are taken against the light have low visibility. In this paper, we propose an image enhancement method to especially improve the visibility of dark areas including subjects in backlit images. First, in the proposed method, dark areas are enhanced by using S-shaped function. Then, the input and enhanced images are fused with weights which are determined by the output of mean or cuboid filter depending on a threshold determined from liner discriminant analysis (LDA). The superiority of the proposed method is demonstrated by qualitative and quantitative comparisons between the proposed method and conventional methods.

14:45–15:00 PCA-based Decolorization Using YPQ Color SpaceAWARD Hayate Shigeoka (Yamaguchi University), Ayaka Fujita (Yamaguchi University), Noriaki Suetake (Yamaguchi University)

Abstract

In this paper, we propose a method for converting digital color images into grayscale images fast and efficiently. In the proposed method, a corrected luminance is obtained by adding values for color correction to luminance value of the original image. Values for color correction are calculated by multiplying color information lost by luminance conversion, scores obtained by performing principal component analysis(PCA) on a chromaticity coordinate, and a parameter. In the experiments, we demonstrate the effectiveness of the proposed method through qualitative and quantitative evaluations.

15:00–15:15 Effects of Particle Size and Motion Speed on Depth Perception in Particle-Based Aerial ImagesAWARD Takumi Mochida (Kindai University), Masanari Asano (Tokyo University of Science), Takanori Koga (Kindai University)

Abstract

This study investigates depth perception in particle-based aerial images. Ten participants observed particle-based sphere and torus images with varying particle size and motion speed. Button responses indicated perceptual switching between flat and depth-enhanced states in all participants, suggesting parameter ranges that may enhance depth perception.

15:15–15:30 Colormap Generation Method Considering Consistency of Color Difference Distributions Between Normal Color Vision and DichromacyAWARD Haruki Hisazaki (Nagoya City University), Go Tanaka (Nagoya City University)

Abstract

Colormaps used for data visualization require appropriate color representation for various color vision types, including dichromacy. In our previous work, we proposed a method to generate colormaps that ensure perceptual uniformity. However, when a user intentionally creates non-uniform color difference patterns, the colormaps need ``perceptual consistency''—which aligns the shape of the color difference distributions between normal color vision and dichromacy—rather than uniformity. In this study, we propose a colormap generation method that replaces the uniformity evaluation term of the previous method with a consistency evaluation term. Experimental results confirmed that our method improves consistency while maintaining contrast in dichromacy for various color combinations.

Slot 2Fri, Oct 30 · 15:45–16:45

[MSS]

Mathematical Systems Science (2)

Fri, Oct 30 15:45–16:45Rooms 2,3 (4F)

Chair: Toshiyuki Miyamoto (Osaka Institute of Technology)

15:45–16:00 Verification of the Phase-Reconstruction Advantage in Periodic Sound Generation Using a Complex-Spectrogram VAE-GANAWARD Masanari Miyaoka (Shibaura Institute of Technology), Kenji Muto (Shibaura Institute of Technology)

Abstract

Herein, we show that phase reconstruction improves the quality of generated periodic cicada sound. Specifically, we developed ComplexSpecVAE-GAN, which treats complex spectrograms as two-channel images and reconstructs waveforms via inverse short-time Fourier transform. Using the sounds of robust cicada (Hyalessa maculaticollis), the proposed framework accurately reproduced waveform periodicity. We added isotropic noise whose magnitude was controlled using a temperature parameter T for robustness analysis. The proposed framework was robust to this noise. However, the noise caused severe performance degradation in the baseline, an amplitude-only variant based on Griffin-Lim, highlighting the structural advantage of phase reconstruction.

16:00–16:15 Proposed-FB: A Full-Bilinear Reward Model for Offline Policy Evaluation in Recommender SystemsAWARD Yuta Masumoto (Waseda University), Manabu Kobayashi (Waseda University), Tota Suko (Waseda University)

Abstract

Proposed-FB is a reward model for recommender offline policy evaluation (OPE). OPE estimates the value of a new policy from logs collected by another policy. Proposed-FB combines feature-derived common representations and identifier (ID)-specific embeddings through a full bilinear matrix. On MovieLens-1M, it achieves the best root mean squared error (RMSE) for reward prediction (0.2231) and direct method (DM) OPE (0.0069), while inverse propensity scoring (IPS) and tree-based doubly robust (DR) results leave its overall OPE advantage preliminary. Recorded dense all-pair evaluation times were 92 ms versus 181 s for LightGBM.

16:15–16:30 Enhancing Matrix Multiplication-free Language Models with Reservoir Computing-inspired LayersAWARD Takumi Shiratsuchi (Kyushu Institute of Technology), Yuichiro Tanaka (Kyushu Institute of Technology), Hakaru Tamukoh (Kyushu Institute of Technology)

Abstract

Although large language models (LLMs) have demonstrated leading performance on natural language processing, their substantial computational demands remain a significant bottleneck. To address the issue, this study focuses on a matrix-multiplication-free language model (MatMul-free LM) and reservoir computing (RC). Concretely, the proposed approach partially freezes and shares the weights of certain layers within the MatMul-free LM, and introduces reservoir recurrent connections that yield adaptive representations at no additional training overhead. Experiments show that the proposed module, a matrix multiplication-free linear gated recurrent unit RC (MLGRU-RC), achieves a 3.9\% reduction in training time and a 6.1\% reduction in inference time, while maintaining performance on par with the baseline model.

16:30–16:45 What Makes Shikaku Puzzles Difficult? A Bipartite Matching ViewAWARD Sosei Kondo (Hiroshima City University), Kazunori Iwata (Hiroshima City University), Takeaki Shimokawa (Hiroshima City University), Kazushi Mimura (Hiroshima City University)

Abstract

This paper proposes a bipartite graph framework for estimating the difficulty of Shikaku puzzles. Blank cells and numbered cells are mapped to two vertex sets, and feasible assignments are represented as edges. Safe edge-reduction rules reduce the matching search space without removing solution-relevant edges. A matching solver records interpretable metrics from propagation, fail-first branching, backtracking, and connected-component decomposition. Experiments on 37 puzzles show that branching decisions increase estimated difficulty, whereas backtracking events have a negative conditional association under the fail-first search strategy. The framework links human difficulty labels to solver execution logs.

[EA]

Engineering Acoustics (2)

Fri, Oct 30 15:45–16:45Rooms 5,6 (4F)

Chair: Yuma Kinoshita (Tokai University)

15:45–16:00 Spectral Envelope Design for Speech-Leakage-Suppressed Audio-Spot Based on Spatial Self-Decoding Using Parametric Array LoudspeakersAWARD Tomoki Hashida (Ritsumeikan University), Yuting Geng (Ritsumeikan University), Takanobu Nishiura (Ritsumeikan University)

Abstract

A parametric array loudspeaker (PAL) can reproduce audible sound in a limited area, called an audio-spot. The separating emission method forms an audio-spot around the intersection of two ultrasonic propagation areas, but speech may leak outside the audio-spot. To address this issue, we propose an audio-spot formation method using spectral envelope design. The target speech is decomposed into two complementary signals and emitted from different PALs. At the audio-spot, the demodulated sounds are spatially combined, enabling spatial self-decoding of the speech envelope. Outside the audio-spot, speech intelligibility is reduced. Subjective evaluation results confirm the effectiveness of the proposed method.

16:00–16:15 Subjective Evaluation of Discomfort Reduction for Residual Noise in Active Noise Control Based on Auditory MaskingAWARD Ryoh Akatsu (Ritsumeikan University), Yuting Geng (Ritsumeikan University), Masato Nakayama (Ritsumeikan University), Yoshiharu Soeta (AIST), Takanobu Nishiura (Ritsumeikan University)

Abstract

Active noise control (ANC) attenuates unwanted noise, but high-frequency peak components may remain in the residual noise because sufficient attenuation is difficult to achieve in basic ANC systems. Such components can cause perceptual discomfort. In this paper, we propose an ANC method that reshapes residual noise by introducing auditory masking into the control process. In the proposed method, a band-stop filter designed according to auditory masking is applied to the error signal, thereby intentionally preserving residual noise around the peak frequency to mask perceptually dominant peak components. Subjective evaluation results showed that discomfort tended to decrease as the expansion width of the critical band was increased.

16:15–16:30 Auditory Masking-Based Crosstalk Suppression for Adjacent Audio Spots Using Parametric Array LoudspeakersAWARD Keiichi Moriyama (Ritsumeikan University), Yuting Geng (Ritsumeikan University), Takanobu Nishiura (Ritsumeikan University)

Abstract

We propose an auditory masking-based crosstalk suppression method for adjacent audio spots using parametric array loudspeakers. The proposed method reshapes the amplitude spectrum of each target signal so that it more effectively masks crosstalk from adjacent spots while preserving its total energy. Specifically, the amplitude correction enhances components where crosstalk is likely to be audible and attenuates components with sufficient masking margin, while penalizing degradation in the quality of the target signals. Evaluation experiments confirmed the effectiveness of the proposed method.

16:30–16:45 Two-stage Point Neuron Network for Direct-Reverberant Sound Field SeparationAWARD Meika Iwamatsu (Tokyo Denki University), Ryosuke Onizawa (Tokyo Denki University), Ayame Uchida (Tokyo Denki University), Izumi Tsunokuni (Tokyo Denki University), Yusuke Ikeda (Tokyo Denki University)

Abstract

We propose a two-stage method based on the Point Neuron Network (PNN) for direct–reverberant sound separation. The PNN satisfies the wave equation through the incorporation of the Green’s function into its network structure, with each neuron corresponding to an equivalent source. In the proposed method, direct and reverberant components are explicitly modeled using equivalent sources. In stage 1, the direct and reverberant components are separated using alternating optimization. In stage 2, the reverberant component is fixed, and the direct-sound equivalent-source positions are refined. Simulation results show that the proposed method outperforms compressed sensing with fixed equivalent-source positions at most frequencies.

[BioX]

Biometrics

Fri, Oct 30 15:45–16:30Room 7 (4F)

Chair: Masatsugu Ichino (The University of Electro-Communications)

15:45–16:00 Preliminary Study on Anonymization of Touch Dynamics Data in Continuous AuthenticationAWARD Riku Hasegawa (The University of Kitakyushu), Yasushi Yamazaki (The University of Kitakyushu)

Abstract

This paper investigates the anonymization of touch dynamics data for continuous authentication. Synthetic impostor data can supplement training data in authentication models, but may retain user-specific characteristics, raising privacy concerns such as re-identification. To address this issue, our previous study generated synthetic data using differential privacy as an anonymization technique. However, anonymization for touch dynamics in continuous authentication has not been sufficiently examined. Therefore, inspired by VAE-based latent representation manipulation for privacy-preserving sensor data, this study applies VAE-based anonymization to flick-operation features and evaluates the resulting data in terms of membership inference resistance, re-identification resistance, and authentication accuracy.

16:00–16:15 A Study on Synthetic Data Generation for Touch Dynamics-Based Continuous AuthenticationAWARD Thanh Phong Do (The University of Kitakyushu), Yasushi Yamazaki (The University of Kitakyushu)

Abstract

This paper presents a new approach for improving the robustness of touch-dynamics-based authentication, including swipe and tap operations on smartphones, as authentication accuracy often degrades due to changes in usage conditions. To address this problem, we propose a new generative model called VAE-GRU-GMM. This model generates diverse synthetic samples while preserving the temporal and spatial characteristics of touch dynamics and reflecting variations in usage conditions. Experiments using a HMOG dataset under sitting and walking conditions showed that incorporating synthetic samples into the training reduced authentication performance degradation, particularly in terms of FRR under walking condition.

16:15–16:30 Personal Authentication Using a Hemispherical Acoustic Ear Recognition Device and a Complex-Input Neural NetworkAWARD Haruki Matsui (Kansai University), Shota Toyooka (Kansai University), Shunsuke Kita (ORIST), Yoshinobu Kajikawa (Kansai University)

Abstract

We propose a hemispherical acoustic ear authentication device that physically mitigates distance variations for personal authentication. Using 20 auricle models, acoustic data were acquired at 121 misaligned locations. To build a robust authentication model, we introduce a novel complex-input 2D convolutional neural network (2DCNN) that treats the real and imaginary parts of complex spectrograms as separate channels. This allows effective learning of both amplitude and phase information. Experimental results show that our method achieves an Equal Error Rate (EER) of 0.56% under positional misalignment. These results indicate that exploiting both amplitude and phase information enables robust acoustic ear authentication against positional misalignment.

Slot 3Sat, Oct 31 · 13:00–14:15

[IT]

Information Theory and Its Applications (1)

Sat, Oct 31 13:00–14:00Rooms 2,3 (4F)

Chair: Hiroki Koga (University of Tsukuba)

13:00–13:15 A Lower Bound on the Adversarial Advantage of Malicious Quantum Adversaries in IKNP OT ExtensionAWARD Muhammad Rusyaidi Zunaidi (Mie University), Akinori Kawachi (Mie University)

Abstract

Oblivious transfer extension reduces the cost of secure computation, yet post-quantum privacy against malicious quantum adversaries is delicate in the quantum random oracle model. This paper derives, for an explicit attack event, a lower bound on the privacy advantage of malicious quantum adversaries against the cut-and-choose IKNP OT extension. The bound combines cut-and-choose survival, Grover-style offending-query search, and a real-ideal statistical-distance argument.

13:15–13:30 Quantum Query Complexity of Approximate Low-Energy CountingAWARD Yuji Shimizu (Mie University), Akinori Kawachi (Mie University)

Abstract

Relative-error approximate low-energy counting, which is the problem of approximately estimating the dimension D of the low-energy subspace defined by a given N-dimensional Hamiltonian with a spectral gap Δ within relative error ε, is a natural quantum analogue of counting solutions to general search problems. We focus on query complexity of this problem in the block-encoding model. We construct an algorithm for relative-error approximate low-energy counting using Õ((1/(εΔ))·√(N/D)) queries to the block-encoding of the Hamiltonian and prove an Ω(ε⁻¹·√(N/D)) lower bound on the number of queries for this problem.

13:30–13:45 Rejection Sampling Using the Konno Density for Quantum-Walk Limit Distributions Ibuki Nakamura (Shizuoka University)

Abstract

This paper proposes a Konno density based rejection sampling method for non-uniform pseudorandom variate generation from 1 dimensional 2 state quantum walk weak limit distributions. These distributions are not Gaussian, and exhibit endpoint singularities. The inverse-CDF method may suffer from numerical integration and interpolation errors near the singular endpoints, while uniform-proposal rejection sampling becomes inefficient due to a low acceptance rate. The proposed method analytically generates candidate samples from the Konno density and accepts them according to the density ratio. Numerical experiments show improved distributional agreement and reduced computation time compared with conventional methods.

13:45–14:00 A Note on Guessing Probability and Min-Entropy for Generalized Symmetric Measurements Masakazu Yoshida (Osaka Sangyo University), Gou Hosoya (Osaka Sangyo University)

Abstract

Uncertainty relations can be viewed as limitations on the predictability of measurement outcomes. We study Landau–Pollak-type uncertainty relations for generalized symmetric measurements in terms of guessing probability and min-entropy. The maximum outcome probability is interpreted as the optimal one-shot guessing probability of the measurement outcome. Under this interpretation, a Landau–Pollak-type uncertainty relation is reformulated as an upper bound on the sum of guessing probabilities, or equivalently on the sum of the exponential forms of the min entropies. We derive bounds for (N, M)-POVMs and discuss examples, including GSIC/SIC POVMs and (d^2−1, 2)-POVMs. In particular, we obtain a tight min-entropy bound for SIC POVMs and show that, for (d^2 − 1, 2)- POVMs, the average guessing probability approaches the uniform binary value as the dimension increases.

[US]

Ultrasonics

Sat, Oct 31 13:00–14:15Rooms 5,6 (4F)

Chairs: Daisuke Koyama (Doshisha University) 13:00–13:30; Shoji Kakio (University of Yamanashi) 13:30–14:15

13:00–13:15 Analysis of thickness-shear mode resonance properties of piezoelectric substrates with periodic voidsAWARD Hikaru Nakajima (University of Yamanashi), Masashi Suzuki (University of Yamanashi), Shoji Kakio (University of Yamanashi)

Abstract

A bonded structure with periodic voids at the interface between a LiNbO3 (LN) thin plate and a support substrate has been shown to exhibit resonance properties similar to those of plate waves in a single LN thin plate. In this study, the thickness-shear mode resonance properties of a bonded structure with periodic voids were analyzed using the finite element method. Two-dimensional analysis investigated the effects of the structure and supporting substrate material. Three-dimensional analysis was conducted to examine the effects of the void configuration. By introducing periodic voids, the resonance properties approached those of a single LN thin plate.

13:15–13:30 Transverse-mode suppression in surface acoustic wave devices on LiNbO3/quartz bonded substrate by proton-exchanged layerAWARD Masahiro Kawahara (University of Yamanashi), Masashi Suzuki (University of Yamanashi), Shoji Kakio (University of Yamanashi)

Abstract

Spurious responses owing to transverse modes appear prominently in high-performance surface acoustic wave (SAW) devices. Although the piston mode approach is a suppression method, challenges remain when applied in high-performance SAW devices, where conventional methods cannot achieve a sufficient velocity difference, affording weak suppression. The proton exchange (PE) layer exhibits the characteristic that acoustic velocity decreases with increasing depth. In this study, a method for suppressing transverse modes using a PE layer was demonstrated through finite element analysis and experiments.

13:30–13:45 Transition from Near-Field to Standing-Wave Acoustic Levitation of a Flat Plate Using a Jumping PhenomenonAWARD Mitsuki Yoshida (Muroran Institute of Technology), Kohei Aono (Muroran Institute of Technology), Hideki Tamura (Tohoku Institute of Technology), Manabu Aoyagi (Muroran Institute of Technology)

Abstract

This study investigates the transition of a flat plate from near-field acoustic levitation to standing-wave acoustic levitation (SWAL) utilizing a jumping phenomenon. Acoustic radiation forces were analyzed by the finite element method, and plate motion was simulated using multibody dynamics. Simulation results showed the plate levitated at approximately 150 μm could be captured and stabilized at a height of approximately 7 mm in SWAL.

13:45–14:00 Single mode vibrator Kazuyuki Ouchi (Wave Device Laboratory)

Abstract

Finite element piezoelectric vibration analysis of quartz oscillators was performed. The electrode geometry was designed according to the vibration mode distribution of the COS structure, and this geometry was applied in the finite element model. The frequency response analysis showed that only the vibration mode corresponding to the electrode pattern was effectively excited. In other words, the excitation of the elastic wave followed the electrode shape distribution. These results demonstrate that the vibration mode of a quartz oscillator can be controlled by appropriately shaping the electrode.

14:00–14:15 Measurement of the Physical Properties of a Droplet Using AM Ultrasound Daisuke Koyama (Doshisha University), Takaya Hirayama (Doshisha University), Kosuke Shigematsu (Doshisha University)

Abstract

Noncontact manipulation technique using ultrasound is expected to be applied to transportation and measurements of physical properties of liquids. In this report, a method to measure the physical property of liquids in air without con-tact using ultrasound was investigated. The experimental system consists of a vibrating disc with four ultrasound transducers and a circular reflector. The generation of an acoustic standing wave between them resulted in the trap-ping of a droplet at the nodal position. The resonant fre-quencies of the droplet depended primarily on the surface tension of the liquids. The fourth modes exhibited reduced variation in measured frequency, indicating that high-er-order modes offer enhanced precision for physical prop-erty measurements.

[SIP]

Signal Processing

Sat, Oct 31 13:00–14:15Room 4 (4F)

Chair: Siqi Na (Nagoya Institute of Technology)

13:00–13:15 Multi-Device RF Identification from Superposed Pilot SignalsAWARD Hanyi Guo (Doshisha University), Tomotaka Kimura (Doshisha University), Jun Cheng (Doshisha University)

Abstract

This paper investigates RF identification of multiple simultaneously active devices from superposed pilot signals. Each device exhibits a unique in-phase/quadrature imbalance, which produces a constellation pattern serving as a discriminative feature. The access point receives a superposition of pilot signals from active devices and forms an estimated constellation pattern. A complex-valued neural network (CvNN) is employed to identify the active device set, where the output classes represent all candidate active-device combinations of size at most K. Simulation results demonstrate that the proposed CvNN outperforms a real-valued neural network in terms of identification error rate.

13:15–13:30 Fast Iterative Shrinkage-Thresholding Algorithm Unfolding for Deep Dictionary LearningAWARD Ataru Shakagori (Fukuoka Institute of Technology), Michiharu Maeda (Fukuoka Institute of Technology)

Abstract

This paper investigates network unfolding and structural constraints on deep k-singular value decomposition (DKSVD) image denoising. Conventional approaches predominantly rely on iterative shrinkage-thresholding algorithm (ISTA) unfolding and sparse coding in dictionary learning has exhibited effective. Our model combines the fast ISTA (FISTA) unfolding with mutual coherence constraints. We examine fast DKSVD (FDKSVD) under this constraint minimizing Gram matrix off-diagonal elements and analyze the dictionary structure through comprehensive experiments on standard datasets. Experimental results demonstrate that our FDKSVD decreases average coherences and improves peak signal-to-noise ratio (PSNR) compared to the conventional DKSVD, optimizing the dictionary structure without compromising the final quality.

13:30–13:45 Illumination-Aware Mode-Switchable Low-Power Automatic Lighting Control System on Board ComputersAWARD Fumiya Ono (Ritsumeikan University), Shunsuke Okura (Ritsumeikan University)

Abstract

To improve smart office energy efficiency, we propose an automatic lighting control system using a compact camera in lighting trims. Unlike conventional motion sensors, image recognition models detect stationary individuals. However, camera-based detection degrades in low light, so the system toggles between two modes based on image brightness: Normal Mode employs segmentation for energy-efficient zone-based LED control, while Dark Mode prioritizes detection via Zero-DCE++ image enhancement. Implemented on a Raspberry Pi Zero 2, the segmentation model achieved a mIoU of 90.6%, with Dark Mode improving low-light accuracy to 99.0%. The system reduces power consumption by an average of 11 W.

13:45–14:00 3.3 - 4.0 GHz -110 dBc/Hz Fractional-N Synthesizer with ADC-Based Direct Phase DigitizationAWARD Thi Viet Ha Nguyen (The University of Electro-Communications), Cong-Kha Pham (The University of Electro-Communications)

Abstract

Fractional-N Sub-sampling Phase-Locked Loop (SSPLL) loop employs a novel analog-to-digital converter (ADC)-based phase detector the need for a digital-to-time converter (DTC). The high PD gain reduces in-band phase noise while its high linearity enables all-digital ΔΣ quantization noise cancellation. The proposed fractional-N SSPLL has been designed in a 180-nm CMOS technology, and the total power consumption was 12.5 mW when a 4-GHz fOUT was generated using a reference frequency of 100 MHz. The RMS jitter and the level of fractional spurs reached 162 fs and -69 dBc, repectively.

14:00–14:15 Adaptive Score-Based VAMP: Self-Tuning Hyperparameters via Tilted EM Siqi Na (Nagoya Institute of Technology), Tadashi Wadayama (Nagoya Institute of Technology)

Abstract

Score-based VAMP (SC-VAMP) provides modular, state-evolution (SE)-predictable inference, but its modules require matched factor hyperparameters; mismatch raises the MSE and degrades matched-SE prediction. We make SC-VAMP adaptive by adding a tilted expectation-maximization (EM) update to each parameterized module, reusing the moments already computed by the SISO interface. Under standard SE population assumptions, the true hyperparameters form a Bayes-optimal population fixed point. Experiments on linear and one-bit Bernoulli-Gaussian compressed sensing show oracle-level performance from strongly mismatched initialization.

[CAS]

Circuits and Systems (1)

Sat, Oct 31 13:00–14:00Room 7 (4F)

Chair: Norihiko Shinomiya (Soka University)

13:00–13:15 The Role of Similar Personal Values and Relationship Ease in Classroom Exclusionary BehaviorAWARD Yoshi Nigorikawa (Niigata University), Takumi Kitamura (Niigata University), Masanori Ori (Niigata University), Kentaro Tani (Niigata University of Health and Welfare), Nao Ito (NIT Toyama College), Kazuyuki Miyakita (Niigata University), Yoshinobu Maeda (Niigata University)

Abstract

The purpose of this study is to investigate classroom bullying and social isolation using a multi-agent simulation model based solely on one-on-one interactions. Agents possess di-verse values and choose behaviors such as Conformity, Inclusion, Distinction, or Exclusion based on dynamically up-dated tolerance thresholds determined by past experiences. The results confirm that a simple and repetitive process of value assessment and threshold updating spontaneously causes group divisions and the structural isolation of specific individuals. This research demonstrates how the accumulation of micro-level interactions drives macro-level group fragmentation.

13:15–13:30 Evaluating a Large Language Model's Perception of Emergent Victimization in an Agent-Based Simulation of Classroom BullyingAWARD Takumi Kitamura (Niigata University), Yoshi Nigorikawa (Niigata University), Kentaro Tani (Niigata University of Health and Welfare), Yoshinobu Maeda (Niigata University)

Abstract

Bullying in classrooms often emerges from ordinary value-laden interactions rather than overt aggression, making victims hard to recognize. In this work, we ask whether a large language model (LLM) can perceive an emergent victim from observable in-formation alone. Coupling an agent-based model of classroom dynamics with an LLM observer, and strictly separating ob-servable inputs from a hidden ground truth, we evaluate per-ception objectively, avoiding circular arguments. Using Lla-ma-3.3-70b, the LLM observer identified acute victims well above chance, but matched long-term emergent victims in only 28% of cases. Furthermore, it exhibited a severe false-positive bias: although baseline snapshots with no active exclusion constituted 60% of all cases, the LLM hallucinated a victim 95% of the time in these instances. Crucially, these tendencies remained invariant across personas, highlighting a robust, state-blind over-detection bias that ignores actual classroom states.

13:30–13:45 Behavior in a Simulation Game for Bullying Reflects Belief in a Just WorldAWARD Masanori Ori (Niigata University), Yoshi Nigorikawa (Niigata University), Tsunehiko Tanaka (Niigata University), Yoshinobu Maeda (Niigata University)

Abstract

Directly observing behavior during real bullying inci-dents is ethically infeasible, limiting behavioral research on bullying. To address this, we developed an Artificial Class Game, set in a virtual classroom, to examine its practical value as a behavioral measurement tool. For 118 fifth- and sixth-grade children, we analyzed the rela-tionship between in-game behavior and belief in a just world (BJW). Exclusion, the only aggressive behavior in the game, correlated negatively with ultimate belief in a just world (BUJ) (r = -.272, p = .003). This game may serve as a valid research tool that elicits behavior re-flecting real psychological traits.

13:45–14:00 An Accuracy Evaluation of Unsupervised Anomalous Sound Detection for Machine Condition Monitoring Using a Spiking Neural Network with Fixed-Point Log-Mel FeaturesAWARD Taishin Hori (Hokkaido University), Chenyu Zhao (Hokkaido University), Hiroshi Tsutsui (Hokkaido University), Takeo Ohgane (Hokkaido University)

Abstract

This paper evaluates the effects of fractional-bit widths used in fixed-point log-mel feature extraction on the detection accuracy of unsupervised anomalous sound detection using a spiking neural network (SNN)-based autoencoder. Sweeping the precision of the Hann window, mel filter bank, and logarithm individually, we identify the minimum fractional-bit width needed at each stage to preserve the accuracy achieved by the floating-point SNN and autoencoder baseline systems. The required minimum fractional-bit width differs greatly among stages, and this ordering is consistent across both systems. The findings guide per-stage fractional-bit allocation to reduce hardware cost.

[SIS]

Smart Info-media System (2)

Sat, Oct 31 13:00–14:15Rooms 201-203 (2F)

Chair: Takanori Koga (Kindai University)

13:00–13:15 Data augmentation for enhanced railway infrastructure component detection under low-light conditionAWARD Natsuki Morita (Tokyo City University), Nico Surantha (Tokyo City University)

Abstract

Aging railway infrastructure requires automated inspection using drones and AI to address labor shortages. However, existing object detection models degrade under adverse weather conditions, and edge devices like the Raspberry Pi have strict computational limits. This study proposes a lightweight detection method using YOLOv7-tiny, trained using a dataset containing clear images together with fog- and low-light-augmented images, without using image enhancement modules. Experimental results demonstrated that the proposed model achieved an mAP@0.5 of 0.953 in adverse conditions, significantly outperforming the baseline clear-weather model (0.587). This approach provides robust, low-computational detection for drone-based automated inspections.

13:15–13:30 Preliminary Software Evaluation of MLP Configurations for ECG Signal Classification on a PYNQ-Z1 PlatformAWARD Jeakwan Kim (Tokyo City University), Nico Surantha (Tokyo City University)

Abstract

Electrocardiogram (ECG) classification is important for automatic cardiac-abnormality analysis, but embedded ECG applications require simple and reproducible models. This paper presents a preliminary software evaluation of multi-layer perceptron (MLP) configurations for ECG beat classification on a PYNQ-Z1 platform. Using the MIT-BIH arrhythmia dataset distributed through Kaggle, three MLP configurations were compared in terms of test accuracy and training time. The optimized MLP achieved 95.10% test accuracy, while the optimized setting with 5-fold cross-validation achieved 97.19% test accuracy and a 99.19% average cross-validation score. The results provide a software baseline for future FPGA-oriented ECG inference.

13:30–13:45 Acoustic Wave Propagation Analysis in Shallow Water for Underwater Acoustic Communication Systems Ryota Goto (Kitami Institute of Technology), Takanobu Kirihara (Kitami Institute of Technology), Shingo Yoshizawa (Chitose Institute of Science and Technology)

Abstract

The range of underwater acoustic communication depends on the amount of energy attenuation associated with sound wave prop-agation, which varies according to water depth and seafloor to-pography. In this study, we present an analysis of acoustic wave propagation reflecting seafloor topography using the parabolic equation (PE). We also evaluate acoustic communication simula-tion in shallow water based on the analysis data.

13:45–14:00 Impulsive Noise Reduction Using Convolutional Neural Network in Underwater Acoustic Communication Reiji Sato (Kitami Institute of Technology), Takanobu Kirihara (Kitami Institute of Technology), Shingo Yoshizawa (Chitose Institute of Science and Technology)

Abstract

In underwater acoustic communication in shallow water, impul-sive noise is frequently observed at high frequencies. The su-perimposition of this noise on the received signal can degrade communication characteristics. Clipping is a method of sup-pressing impulsive noise, however, an appropriate threshold must be set according to the amplitude of the communication signal. In this study, we propose an impulsive noise removal method using a convolutional neural network (CNN) that is in-dependent of threshold values. We report the simulation results on communication performance.

14:00–14:15 Frequency-Domain Narrowband Denoising Autoencoder in Underwater Acoustic Communication Systems Yuki Utashiro (Kitami Institute of Technology), Takanobu Kirihara (Kitami Institute of Technology), Shingo Yoshizawa (Chitose Institute of Science and Technology)

Abstract

In underwater acoustic communications, narrowband noise gen-erated by fish finders and underwater robots can degrade de-modulation performance. This study proposes using a frequen-cy-domain denoising autoencoder to remove narrowband noise. The frequency spectrum is extracted from input data in which narrowband noise is superimposed on the communication signal. Learning and noise removal are performed using a denoising autoencoder. We evaluate the communication characteristics through simulations and actual measurements.

[ISEC]

Information Security (1)

Sat, Oct 31 13:00–14:15Hall (2F)

Chair: Koki Jimbo (Tokyo University of Science)

13:00–13:15 Query Recovery Attacks Utilizing Volume and Search Patterns against Searchable EncryptionAWARD Ryousuke Muramatsu (Future University Hakodate), Takumi Namiki (The University of Electro-Communications), Yohei Watanabe (The University of Electro-Communications)

Abstract

Leakage abuse attacks are a line of research that carefully analyzes the practical impact of the leakage during operations in searchable encryption. In this paper, we improve SimGraphVL, one of the state-of-the-art leakage abuse attacks proposed by Namiki et al. (ESORICS 2025), and propose a new leakage abuse attack called SimGraph+VL. We demonstrate experimentally that our attack recovers more search queries than SimGraphVL.

13:15–13:30 Improving the Communication Efficiency of MPC-in-the-Head Zero-Knowledge Proofs via Fiat--Shamir TransformationAWARD Yuto Goto (The University of Electro-Communications), Mitsugu Iwamoto (The University of Electro-Communications), Yohei Watanabe (The University of Electro-Communications)

Abstract

Hazay et al. proposed zero-knowledge proof protocols extending MPC-in-the-Head using game-based cryptographic primitives. Its communication does not directly rely on its circuit size. We reduced the communication complexity of HSS-based ZKP, one of Hazay’s protocols, by reducing the interaction in randomness generation. Specifically, we replaced interactive random numbers with a pseudorandom function whose input is the prover's randomness. This achieves reducing communication size, but fails to satisfy the zero knowledge property. By applying the Fiat-Shamir transformation, we obtain full zero knowledge in the random oracle model while preserving the communication improvement.

13:30–13:45 Strategic Security of Dynamic Collateral in Decentralized Sealed-Bid AuctionsAWARD Kholiq Abdullah (The University of Electro-Communications), Yohei Watanabe (The University of Electro-Communications)

Abstract

Decentralized sealed-bid auctions provide bid confidentiality, correctness, and public verifiability without a trusted auctioneer. Existing security models mainly consider a single auction with static collateral. In concurrent auctions, however, bidders may share collateral across multiple instances, introducing threats such as collateral overcommitment, information leakage, inconsistent claims, and proof reuse across different auction contexts. This paper models dynamic collateral management as a cryptographic state consistency problem and identifies four requirements: solvency safety, collateral privacy, collateral integrity, and cross-auction consistency. Rather than analyzing gametheoretic equilibria, we focus on cryptographic mechanisms that prevent strategic collateral misuse.

13:45–14:00 Efficient GGM-Based Pseudorandom Functions Using a d-ary Tree and the BBS PRGAWARD Aya Hakamata (The University of Electro-Communications), Kazuha Murakami (The University of Electro-Communications), Hayato Gibo (The University of Electro-Communications), Kazuki Sato (The University of Electro-Communications), Yohei Watanabe (The University of Electro-Communications), Mitsugu Iwamoto (The University of Electro-Communications)

Abstract

We study an efficient construction of pseudorandom functions (PRFs) based on the GGM tree. We combine a d-ary GGM tree with the Blum–Blum–Shub pseudorandom generator (BBS PRG) and propose a method that skips unnecessary output bits. We analyze the evaluation cost of the resulting PRF and show that the proposed construction achieves a lower actual evaluation time than the best basic d-ary GGM construction with sequential BBS PRG generation.

14:00–14:15 DLT Terms 100: a learner's glossary based on ISO 22739:2024 Yusuke Kaneko (The Japan Research Institute), Takumi Aita (The Japan Research Institute), Kohei Ichihara (The Japan Research Institute), Hiroki Watanabe (The Japan Research Institute)

Abstract

We propose ``DLT Terms 100'' aimed at improving literacy regarding blockchain (distributed ledger technology: DLT). The pace of DLT's social implementation has been slower than initially expected, partly due to a shortage of engineers and literate individuals. In particular, the discrepancy between business and technical terminology, and the confusion caused by the repurposing of existing financial terms, are barriers to understanding the concepts and technologies. Our proposed glossary is based on ISO 22739:2024, and categorizes and labels terms based on their frequency of appearance in representative domestic and international literature and their difficulty level.

Slot 4Sat, Oct 31 · 14:30–15:45

[IT]

Information Theory and Its Applications (2)

Sat, Oct 31 14:30–15:30Rooms 2,3 (4F)

Chair: Akinori Kawachi (Mie University)

14:30–14:45 On the Equivalence of Two Coded Caching Schemes with Heterogeneous Cache SizesAWARD Tomoki Yoneyama (University of Tsukuba), Hiroki Koga (University of Tsukuba)

Abstract

In the coded caching scheme proposed by Maddah-Ali and Niesen [1], we consider the situation where K users have respective cache memories and send requests for one of the N files to a server. Coded caching schemes with heterogeneous cache sizes are discussed by several authors. Recently, Nagaya et al. [2] and Sengupta et al. [4] proposed the coded caching schemes for heterogeneous cache sizes. In this paper, we theoretically analyze the transmission rates of the schemes in [2] and [4] and show that the optimum rate of the two schemes coincide.

14:45–15:00 Threshold Compensation for Recovering Absolute Capacity in Dense Associative Memory with Biased PatternsAWARD Yuto Sakurai (Hiroshima City University), Takeaki Shimokawa (Hiroshima City University), Kazunori Iwata (Hiroshima City University), Kazushi Mimura (Hiroshima City University)

Abstract

Dense associative memory stores many patterns as stable states of recurrent networks. We study centered biased patterns whose entries are 1-q with probability q and -q with probability 1-q. Centering removes the pattern mean, but biased symbols create an imbalance in the local decision rule. We introduce threshold compensation and derive leading-order absolute capacity estimates with and without it. Numerical experiments for n=3 show that threshold compensation improves finite-size capacity estimates.

15:00–15:15 A Note on Performance Evaluation of CM Capacity-Achieving DBICM Systems for APSK Constellations by Density Evolution Gou Hosoya (Osaka Sangyo University)

Abstract

Delayed bit-interleaved coded modulation (DBICM) reorders part of the coded bits in time, so that bits already received help to decode later ones. For APSK its middle-region capacity reaches the coded-modulation (CM) capacity at a maximum delay of m − 2. This note evaluates such a system for a 16APSK(4+12) constellation by density evolution. We design the boundary pilots from the CM capacity and compare delay vectors of different maximum delay dₘₐₓ. At a matched code rate the dₘₐₓ = 2 and staircase (dₘₐₓ = 3) vectors perform equally, and every vector gains about 1.6 dB over a rate-matched BICM baseline.

15:15–15:30 On the Asymptotic Competitive Optimality of Bayes Codes Shota Saito (Gunma University)

Abstract

This paper considers the problem of universal lossless source coding. The Bayes code is one of the universal lossless source codes, which is designed so that it minimizes the Bayes average codeword length. We demonstrate that the Bayes code is not only optimal in the average sense but also asymptotically competitively optimal for a context tree source.

[EA]

Engineering Acoustics (3)

Sat, Oct 31 14:30–15:30Rooms 5,6 (4F)

Chair: Toshiyuki Kimura (Tohoku Gakuin University)

14:30–14:45 2D Sound Field Interpolation Using Physics-Informed Extreme Learning Machine with Pre-trainingAWARD Hayato Komaba (Tokyo Denki University), Gen Sato (Tokyo Denki University), Yusuke Ikeda (Tokyo Denki University)

Abstract

Sound field interpolation estimates sound pressure distributions at unobserved locations and has been widely applied to acoustic applications requiring 2D processing. This paper proposes a Physics-informed extreme learning machine (PIELM)-based 2D sound field interpolation method. Hidden-layer parameters are pre-trained using a Physics-informed neural network (PINN), and output-layer parameters are obtained through matrix operations. Experimental results demonstrate that the proposed method significantly accelerates training in 2D sound field interpolation tasks requiring an NMSE of approximately -10 dB.

14:45–15:00 A Remote Microphone-based Virtual Sensing Method Using Two Compensation Filters in Active Noise ControlAWARD Shintaro Kada (Kansai University), Shota Toyooka (Kansai University), Kenta Iwai (Osaka Sangyo University), Yoshinobu Kajikawa (Kansai University)

Abstract

This paper proposes a novel remote microphone-based virtual sensing method employing two compensation filters to enhance noise reduction in frequency bands where conventional approaches provide insufficient attenuation. The proposed method first estimates the coherence between the virtual microphone and two error microphones. For each frequency bin, only signal components from the error microphone with higher magnitude-squared coherence are used to update the noise control filter, thereby improving the reliability of the adaptation process. Numerical simulations demonstrate that the proposed method achieves improved noise attenuation in frequency bands that are difficult to attenuate using conventional remote microphone techniques.

15:00–15:15 Preliminary Study on a Frequency-Domain ANC System Incorporating Phase-Error Compensation for the Secondary-PathAWARD Haruto Kitakusu (Kansai University), Shota Toyooka (Kansai University), Kenta Iwai (Osaka Sangyo University), Yoshinobu Kajikawa (Kansai University)

Abstract

This paper proposes a frequency-domain active noise control (ANC) system incorporating step-size sign control to suppress instability caused by secondary-path modeling errors. In conventional Filtered-x algorithms, which are widely used in ANC systems, a phase error greater than 90° between the actual secondary path and its model can result in an inappropriate update direction of the adaptive filter, thereby causing system instability. In the proposed system, the phase error between the secondary path and its model is measured in advance. A negative step size is then assigned to frequency bins where the phase error exceeds 90°, thereby correcting the update direction of the adaptive filter and preventing divergence due to secondary-path modeling errors. Simulation results demonstrate that the proposed system remains stable even under conditions in which the conventional system diverges, achieving a noise reduction of approximately 10 dB. These results indicate that the proposed system is effective in improving robustness against secondary-path modeling errors.

15:15–15:30 Investigation of Discrimination of Pitch and Tone-Height for Voice Shota Morita (Fukuyama University), Teruki Toya (University of Yamanashi)

Abstract

In this paper, we investigate the pitch and tone-height discrimination characteristics of Japanese vowels, nasals, and words by systematically varying the fundamental frequency and spectral envelope of self-uttered and other-uttered voice. Participants compared reference stimulus with a comparison stimulus and judged whether the latter was lower, the same, or higher than the former. Correct response rates were higher for downward than for upward changes in both pitch and tone-height discrimination. Furthermore, for fundamental frequency changes exceeding ±4 Hz, voice stimuli exhibited discrimination characteristics similar to those of pure and harmonic complex tones.

[IMQ]

Image Media Quality

Sat, Oct 31 14:30–15:15Room 4 (4F)

Chair: Masaru Tsuchida (NTT, Inc.)

14:30–14:45 Color Image Enhancement with Mean Intensity and Mean Saturation Preservation Based on Histogram SpecificationAWARD Ayato Isaida (Tokyo City University), Akira Taguchi (Chitose Institute of Science and Technology)

Abstract

This paper proposes a universal color image enhancement method that simultaneously improves contrast and colorfulness while strictly preserving the mean intensity and mean saturation of the input image. However, Jiang et al.'s conventional histogram specification method is restricted to a narrow range of mean values. To overcome this, we introduce novel target histogram models combining rectangular and triangular shapes for low and high mean value regions. Applying this approach independently to intensity and saturation components in the HSI color space enables natural enhancement without color shifting. Experimental results using 26,664 images demonstrate the effectiveness of our method.

14:45–15:00 Equalization of Signal Distribution in Color Space under Hue PreservationAWARD Takeshi Goto (Tokyo City University), Akira Taguchi (Chitose Institute of Science and Technology)

Abstract

Images captured under poor illumination conditions always exhibit low contrast and colorfulness. This study aims to improve these deficiencies by equalizing the signal distribution in the HSI color space. To achieve this, target histograms for intensity and saturation are derived based on the geometric characteristics of iso-hue planes, and histogram specification is applied to equalize the signal distribution in the color space. Subjective and objective evaluations demonstrate that the proposed method effectively enhances both contrast and colorfulness.

15:00–15:15 Analysis and Optimization of Background Textures in Capture-Resistant Data HidingAWARD Kyosuke Inoue (National Institute of Technology, Tokyo College), Tetsuya Kojima (National Institute of Technology, Tokyo College)

Abstract

In this paper, we analyze the factors causing data extraction failure in data hiding techniques for printed media during the capture process and derive the optimal conditions for accurate extraction. We evaluate how background textures affect extraction accuracy from the viewpoint of the geometric correction of captured images and the host signal interference. It is shown that the geometric correction fails due to a lack of feature points and incorrect matching. Regarding host signal interference, it is also shown that bit errors occur when the power of the background’s frequency components exceeds that of the embedded signal. Experimental results reveal the optimal embedding strength that simultaneously ensures the stability of the geometric correction and suppresses host signal interference.

[CAS]

Circuits and Systems (2)

Sat, Oct 31 14:30–15:15Room 7 (4F)

Chair: Yoshinobu Maeda (Niigata University)

14:30–14:45 Enhancing Memory Capacity of Reservoir Computing with Delayed Reservoir States for FPGA ImplementationAWARD Tomoya Hirano (Kyushu Institute of Technology), Kanta Yoshioka (Kyushu Institute of Technology), Soshi Hirayae (Kyushu Institute of Technology), Hakaru Tamukoh (Kyushu Institute of Technology)

Abstract

Reservoir computing is a promising approach for low-power, real-time time-series processing on edge devices. Look-up table network-based reservoir computing (LUTNet-RC) implements reservoir neurons using LUT elements on an field-programmable gate array with low latency. However, its memory capacity which measures the ability to retain past input information is often insufficient for tasks that require long-term temporal dependencies. This paper proposes a delayed-output layer structure that stores past reservoir states and integrates them into output layer. Although the proposed method roughly doubles circuit resource usage, experimental results on benchmark tasks show that it improves memory capacity by up to 7.8 times.

14:45–15:00 FPGA Streaming Integration and 30-fps Feasibility Evaluation of a Deep-Learning-Free Gesture Recognizer Based on LUTNet Reservoir ComputingAWARD Tetsuya Hayashi (Kyushu Institute of Technology), Kanta Yoshioka (Kyushu Institute of Technology), Soshi Hirayae (Kyushu Institute of Technology), Hakaru Tamukoh (Kyushu Institute of Technology), Yuichiro Tanaka (Kyushu Institute of Technology)

Abstract

Low-power always-on gesture recognition is essential for service robots. Although prior work demonstrated deep-learning-free recognition with look-up table network reservoir computing (LUTNet-RC), the feasibility of register-transfer-level (RTL) field-programmable gate array (FPGA) implementation had not been verified. Toward FPGA integration, this paper presents an RTL-based streaming integration core connecting preprocessing and inference for binary frame-difference input images. Post-route implementation on an Artix-7 device meets 100-MHz timing using 9.53\% LUTs and no digital signal processing (DSP) blocks. RTL simulation processes a 640×480 frame in 1.548 ms and matches a fixed-point Python reference (85.4\% video-level classification accuracy). On-board verification remains future work.

15:00–15:15 Pseudo-Differential Cyclic ADC with Relaxed Requirements on Analog ComponentsAWARD Yusuke Morishita (Tokyo City University), Hao San (Tokyo City University), Masao Hotta (Tokyo City University), Tatsuji Matsuura (Tokyo City University)

Abstract

Cyclic analog-to-digital converters (ADCs) are attractive for IoT sensors, monitoring and control systems owing to their high area efficiency. Nevertheless, conventional cyclic ADC designs rely on fully differential high gain amplifiers, meticulous frequency characteristics, and common-mode feed-back (CMFB) circuits, which increase circuit overhead. To mitigate these issues, this paper presents a pseudo-differential cyclic ADC architecture with a simplified amplifier without CMFB. The proposed ADC was designed in 65nm CMOS technology, and simulations show an SNDR of 65.98 dB and an ENOB of 10.67 bits, confirming the feasibility of cyclic ADC operation with relaxed amplifier complexity.

[SIS]

Smart Info-media System (3)

Sat, Oct 31 14:30–15:30Rooms 201-203 (2F)

Chair: Yosuke Sugiura (Meiji University)

14:30–14:45 Understanding of Walking Environment Employing MYVISION - A Method of Detecting StairsAWARD Ryusei Nakashima (Kyushu Institute of Technology), Yui Tanjo (Kyushu Institute of Technology)

Abstract

Reliable stairs detection is essential for visually impaired individuals. This paper proposes a method of DBSCAN-based clustering and dataset enhancement using non-stair images for an RGB-D image-based stair detection. Experimental results show that the proposed method improved the F1-score from 0.51 to 0.63 and reduced false detections cases in non-stair environments, while revealing a trade-off between false detection suppression and stair detection performance.

14:45–15:00 Low-Light Image Enhancement Based on Convex Combination Representation of Colors in CIELAB Color SpaceAWARD Keita Murayama (Ryukoku University), Yoshiaki Ueda (Ryukoku University)

Abstract

Low-light image enhancement is a challenging task that aims to recover perceptually meaningful information from low contrast images with severely degraded visual information. In this study, we propose a low-light image enhancement method that focuses on accurate hue restoration by representing pixels as convex combinations in the CIELAB color space. Experimental results show that the proposed method achieves improved hue restoration accuracy compared with conventional methods.

15:00–15:15 A Two-Stage Framework with a TFC-TDF-UNet-Based Second Stage for Six-Stem Music Source SeparationAWARD Juzheng Lin (Hokkaido University), Hiroshi Tsutsui (Hokkaido University), Takeo Ohgane (Hokkaido University)

Abstract

Conventional music source separation performs well under the four-stem setup. However, research on extended sources remains limited, restricting its applicability to practical scenarios. To extend separation to six stems and improve the quality of guitar and piano, we propose a two-stage separation framework. The first stage separates vocals, drums, and bass, while the second stage further extracts guitar and piano from the residual signal using TFC-TDF-UNet-based methods. This cascaded approach can leverage the strengths of pretrained models while reducing the difficulty of separating instruments with complex spectra. Our system achieves competitive performance when evaluated on the MoisesDB dataset.

15:15–15:30 A Time Synchronization Method of a Self-Measurable Electronic Stopwatch for Sprinting Using a Bluetooth-Connected Single-Board Computers Hiroaki Hata (Tottori University of Environmental Studies), Kazuhiro Matsui (NIT Fukui College), Akihiro Azuma (NIT Fukui College)

Abstract

We are developing a stopwatch that a runner can use to time themselves for sprint training. This system for a single sprint-er uses two small computers: one records the start signal and time, while the other detects the finish with a photoelectric sensor. The devices communicate via BLE to exchange tim-ing data and calculate elapsed time, but BLE alone cannot meet the required synchronization accuracy of within 10 milliseconds. To solve this, the computers are physically connected through GPIO pins only for time synchronization, controlled by BLE. After synchronization, they are discon-nected and used for the timing measurement.

[ISEC]

Information Security (2)

Sat, Oct 31 14:30–15:45Hall (2F)

Chair: Yohei Watanabe (The University of Electro-Communications)

14:30–14:45 Computational Cost of Asymmetric Diffie-Hellman Key AgreementAWARD Kento Murota (Tokyo University of Science), Masanari Asano (Tokyo University of Science), Koki Jimbo (Tokyo University of Science)

Abstract

This study proposes a model that evaluates the computational costs of an asymmetric elliptic-curve Diffie-Hellman protocol for parties with resource constraints. The goal is to demonstrate the usefulness of the protocol. This model explains the relationship between the degree of a protocol’s asymmetry and its computational time which is consistent with the measurement results obtained from devices under given communication environments.

14:45–15:00 A Known-Plaintext Key-Recovery Attack on the RCS Toy Model Ryuto Ishii (AMIYA corporation), Koki Jimbo (Tokyo University of Science), Maki Kihara (Konan University)

Abstract

The RNA-Crypto System (RCS) is a symmetric-key cryptographic algorithm inspired by RNA splicing, where ciphertext blocks are interpreted as exons carrying plaintext information or introns carrying dummy information. In this paper, we analyze the RCS toy model in a known-plaintext setting. We construct a key-recovery attack that uses a plaintext-ciphertext pair to derive candidates for each block key. The attack exploits the 1-bit-shift rule of block key generation, which makes adjacent block keys overlap in two bits. This overlap eliminates impossible exon-intron sequences and reduces the number of secret-key candidates.

15:00–15:15 Bootstrapping Reduction for Homomorphic Keccak256 over TFHEAWARD Yuto Kitadai (Okayama University), Yasuyuki Nogami (Okayama University), Yuta Kodera (Okayama University)

Abstract

Fast Fully Homomorphic Encryption over the Torus (TFHE) supports efficient Boolean evaluation on encrypted data. This makes it attractive for bit-oriented primitives such as Keccak-256 in confidential smart contracts. However, TFHE gate evaluation relies heavily on bootstrapping, the dominant performance bottleneck. Hong et al. mitigated the cost in Keccak-256 evaluation via overlapped bootstrapping, achieving a 42--44\% speedup. A substantial cost nevertheless remains. This work applies output-cone pruning to the final round of Hong et al.'s model55. Experiments confirm a further runtime reduction of 3.3\%, or 75 seconds. This brings homomorphic Keccak-256 evaluation closer to practical use in confidential smart contracts.

15:15–15:30 A Preliminary Implementation and Evaluation of Privacy-Preserving Federated Learning Platform using CDKS.Dec with SSSAWARD Takumi Tojo (Meijo University), Ryo Kumagai (Meijo University), Shu Takemoto (Meijo University), Yusuke Nozaki (Meijo University), Masaya Yoshikawa (Meijo University)

Abstract

Privacy-preserving federated learning (PPFL) is required to keep local models secret from the server. CDKS, a multi-key homomorphic encryption scheme, enables PPFL but has a vulnerability allowing individual models to be recovered during decryption. CDKS.Dec with SSS addresses this using Shamir's secret sharing, yet remains a theoretical proposal. This study implements CDKS.Dec with SSS in federated learning, featuring fixed-point weight encoding and dropout handling. Experiments confirm that training is correctly established under encryption and remains effective under client dropout.

15:30–15:45 Experimental Evaluation of Colored QR Codes that Output Two Types of Information Depending on the Shooting Environment Masanori Hirotomo (Saga University), Kaiya Numata (Saga University), Makoto Takita (Kobe University), Yoshiaki Shiraishi (Kobe University)

Abstract

In this paper, we experimentally evaluate QR codes that output two types of information depending on the shooting environment, for the case where the dark module is a color other than black. The results show that the decoding result changes under a weak-light environment (in the shade), an opposite tendency to the previous work, where it changes under strong light. This demonstrates the existence of an environment-dependent passive attack even for QR codes using corporate colors.

Slot 5Sat, Oct 31 · 16:00–17:15

[IT]

Information Theory and Its Applications (3)

Sat, Oct 31 16:00–16:45Rooms 2,3 (4F)

Chair: Shota Saito (Gunma University)

16:00–16:15 Experimental evaluation of MPC attack against Secret Sharing with SnitchingAWARD Ryota Fujimori (Mie University), Akinori Kawachi (Mie University)

Abstract

In secret-sharing schemes, if some shareholders maliciously collude, the shared secret may be illegally reconstructed. Moreover, detecting and punishing such misbehavior is generally difficult. To address this problem, Secret Sharing with Snitching (SSS) has been proposed. Prior work suggests that the computation-time gap between ordinary hash evaluation and multiparty computation (MPC) is related to the security of SSS. However, this gap has not been experimentally clarified under realistic parameters. Therefore, in this work, we measured execution times with and without MPC to evaluate the practicality of SSS. The results show that using MPC is approximately 3.0 × 10^5 times slower.

16:15–16:30 Converse Bounds for SPIR with User-Side Common Randomness Under Limited Information LeakageAWARD Yuta Watanabe (Saitama University), Tetsunao Matsuta (Saitama University)

Abstract

Private information retrieval (PIR) allows a user to download a desired message from multiple servers without revealing which message is requested. Symmetric PIR (SPIR) additionally prevents information leakage about undesired messages to the user. This paper considers SPIR with user-side common randomness and limited leakage about undesired messages. For this model, we give converse bounds on the normalized download cost and the normalized amounts of server-side and user-side common randomness.

16:30–16:45 On the Optimal Relative Difference of Superimposed All the Shares for a (3,n)-Threshold Visual Cryptography Hiroki Koga (University of Tsukuba), Naoki Mitarai (University of Tsukuba)

Abstract

A (t, n)-threshold visual cryptography scheme ((t, n)-VCS for short) provides a nonconventional secret sharing scheme for digital black-white images. In the (t, n)-VCS, a dealer generates n shares from a secret image, prints each share on a transparency sheet, and distributes the printed n shares to the respective n participants. While arbitrary t participants can recover the secret image by superimposition of their shares, no information on the secret image is revealed from any less than t shares. In this paper, we focus on the total relative difference αₙ that represents the contrast between black and white pixels in the image obtained by superimposition of all the n shares. We succeed in obtaining a tight upper bound of αₙ for the case of t = 3 by using a combinatorial argument. This upper bound leads to the optimal construction of the (3, n)-VCS using a BIBD.

[EA]

Engineering Acoustics (4)

Sat, Oct 31 16:00–17:00Rooms 5,6 (4F)

Chair: Takanobu Nishiura (Ritsumeikan University)

16:00–16:15 Three-Dimensional Acoustic Playback Evaluation Method Using Music SignalAWARD Yui Yashima (Tokyo University of Science), Benedikt Bugl (OTH Regensburg), Viola Isabella Schneider (OTH Regensburg), Gerald Schickhuber (OTH Regensburg), Armin Sehr (OTH Regensburg), Takahiro Yoshida (Tokyo University of Science)

Abstract

In our previous research, a three-dimensional acoustic play-back evaluation method, using a time-stretched pulse as a test tone and recording the signal using an Ambisonics micro-phone, has been developed. This method enables the visual-ization of the acoustic playback condition at the listening position. This study proposes a method that uses musical signals as test tones to enable analysis during music playback. In evaluations of stereo loudspeaker misalignment, we con-firmed that our method can successfully analyze it with performance similar to the previous method using the TSP signal. Moreover, we evaluated the performance using five musical genres in three different rooms.

16:15–16:30 Sound Field Transformation Across Different Boundary Conditions Using a Single Room Impulse ResponseAWARD Itsuki Theo Terashita (Tokyo Denki University), Gen Sato (Tokyo Denki University), Yusuke Ikeda (Tokyo Denki University)

Abstract

We propose a deep learning method for transforming sound fields between different　boundary conditions using a single room impulse response (RIR) as a conditioning signal. A conditioning vector is generated from the log-amplitude spectrum of the observed RIR and used to modulate a U-Net via Featurewise Linear Modulation (FiLM).Experiments on　simulated 2D room sound fields yield a full-field SNR of 24.2 dB and a reflected-field SNR of 14.8 dB, demonstrating that a single RIR serves as an effective conditioning signal for sound field transformation.

16:30–16:45 Relation-Network-Based 3D Position Estimation for Variable Numbers of Microphone-Embedded LoudspeakersAWARD Chihiro Nishihara (Tokyo Denki University), Gen Sato (Tokyo Denki University), Izumi Tsunokuni (Tokyo Denki University), Yusuke Ikeda (Tokyo Denki University)

Abstract

Smart speakers integrating microphones and loudspeakers are promising devices for sound field reproduction. Accurate loudspeaker positions are essential but difficult to obtain in real environments. In addition, the number of available speakers may vary depending on the environment. Our previous work proposed a Relation Network (RelNet)-based deep learning method using a single model for varying numbers of speakers; however, it was limited to two-dimensional localization. In this study, we extend this RelNet-based method to three-dimensional speaker position estimation. Simulation results showed that, although the proposed method was less accurate than a fixed-unit model, it successfully estimated 3D positions for different numbers of speakers using a single model.

16:45–17:00 Theoretical Study of Scalable 3D Radiated Sound Field Display System Using Directional Loudspeakers and Wave Field Synthesis Toshiyuki Kimura (Tohoku Gakuin University)

Abstract

In this paper, I theoretically study the sound field reproduced in the previously proposed three-dimensional sound field reproduction system when the size of the surrounding microphone array is different from that of the radiated loudspeaker array. I show that the three-dimensional radiating sound field, in which the speed of sound changes and the size is expanded or contracted, is faithfully reproduced when the sound recorded by the surrounding microphone array is reproduced by the radiated loudspeaker array, if the arrangement of the microphones and directional loudspeakers is similar.

[VLD]

VLSI Design Technologies

Sat, Oct 31 16:00–16:45Room 4 (4F)

Chair: Toshiyuki Ichiba (Fujitsu)

16:00–16:15 Investigation of the Applicability of Approximate Arithmetic Units to Artificial Vision Systems for Power ReductionAWARD Saishi Taira (Kindai University), Tomoaki Ukezono (Kindai University)

Abstract

This paper evaluates the applicability of approximate adders to the subtraction operation in the Exponential-Golomb decoding process in an artificial vision system as a preliminary step toward reducing power consumption. Stimulus pattern data representing binary visual stimulation patterns encoded using Run-Length Encoding Exponential-Golomb Coding (RLE-EGC) were decoded with the subtraction operation implemented using nine 16-bit signed approximate adders from EvoApproxLib. The reconstructed images were evaluated using PSNR and SSIM. Experimental results showed that all evaluated approximate adders degraded image quality, indicating that approximation errors cannot be neglected in Exponential-Golomb decoding.

16:15–16:30 Design of a Single-Port SRAM-Based Configurable-Length FFT IP Core with Four-Cycle Butterfly Operations Using a Single MultiplierAWARD Shota Ando (Hokkaido University), Chenyu Zhao (Hokkaido University), Hiroshi Tsutsui (Hokkaido University), Takeo Ohgane (Hokkaido University)

Abstract

This paper presents a small-scale, low-power fast Fourier transform (FFT) IP core using a single-port SRAM. Supporting configurable power-of-two transform lengths, the architecture processes each butterfly operation over four clock cycles by reusing a single multiplier. The synthesis results show that the proposed IP core reduces the NAND gate count of the complex multiplier by about 75 %, achieving an overall reduction of 24.8 % for a 1,024-point FFT compared with a baseline module executing butterfly operations in one cycle.

16:30–16:45 Case Study of Coding-Agent Optimization by Synthesis-in-the-Loop for Crypto-Accelerator DesignAWARD Soshi Tsunashima (The University of Tokyo), Makoto Ikeda (The University of Tokyo)

Abstract

We study how an LLM coding agent optimizes arithmetic-oriented cryptographic hardware, where arithmetic algorithms and datapath organization, not control logic, decide quality, when logic-synthesis reports are fed back into its design loop. Under a fixed, reproducible protocol, a commercial coding agent designed ECDSA P-256 verification circuits in two eight-revision trajectories that differ only in where the area–time objective is stated. Both reduced the area–time product by more than 4×. The agent predicted cycle counts within ±1.2% but systematically mispredicted synthesis outcomes; it freely revised arithmetic algorithms yet never revisited its initial datapath organization; and its progress was non-monotone. We discuss implications for evaluating coding agents on arithmetic-oriented hardware.

[WBS]

Wideband System

Sat, Oct 31 16:00–17:00Room 7 (4F)

Chair: Taishi Watanabe (KDDI Research, Inc.)

16:00–16:15 Throughput Analysis of Threshold-Aided Two-Device BSIC-SAAWARD Ryo Ozaki (Doshisha University), Daiki Fukui (Doshisha University), Yuhei Takahashi (Doshisha University), Tomotaka Kimura (Doshisha University), Jun Cheng (Doshisha University)

Abstract

We study threshold-aided buffered-SIC (successive interference cancellation) slotted ALOHA, referred to as BSIC-TA-SA, over Rayleigh fading channels. In the proposed scheme, each device transmits with a given probability only when its instantaneous received SNR exceeds a decoding threshold, thereby suppressing low-SNR transmissions and reducing interference. For a two-device system, state transition probabilities are derived for a Markov process, enabling an explicit throughput expression. The transmission probability is optimized for a given spectral efficiency. Numerical and simulation results demonstrate the throughput improvement of the proposed scheme and validate the analysis.

16:15–16:30 Performance Evaluation of Adaptive Control using Water-Filling in Acoustic PORC-OFDM SystemsAWARD Mahiro Munakata (Shinshu University), Fumihito Sasamori (Shinshu University), Ryo Shibata (Shinshu University)

Abstract

Acoustic communication, which uses sound waves as the transmission medium, suffers from low spectral efficiency because it operates over frequency-selective channels. The conventional adaptive Phase-Offset Repetition-Coded OFDM (PORC-OFDM) scheme improves spectral efficiency by adaptively controlling both the modulation order and the repetition factor according to the channel condition of each subcarrier. To further improve spectral efficiency, this paper proposes an enhanced adaptive PORC-OFDM scheme incorporating a water-filling-based power allocation method. Acoustic communication experiments demonstrate that the proposed scheme achieves higher throughput than the conventional scheme over the evaluated range of transmission distances.

16:30–16:45 Rapid On-site 3D Measurement Support via MASt3R-SLAM and Depth-Regularized Gaussian SplattingAWARD Kaito Kusano (Aichi Prefectural University), Haruki Kawanaka (Aichi Prefectural University), Koji Oguri (Aichi Prefectural University)

Abstract

We propose a fast smartphone-based framework for 3D measurement support that enables rapid on-site visualization to ensure data completeness. Conventional 3D Gaussian Splatting (3DGS) commonly relies on time-consuming Structure-from-Motion (SfM) for initialization, which hinders immediate feedback in the field. We instead leverage MASt3R-SLAM, a learning-based dense SLAM system, to provide camera poses and geometry. By converting its output to COLMAP-compatible records, our pipeline reduces initialization time by 84.0\% and total processing time by 47.1\% compared with a COLMAP-based reference. The approximately 14-minute turnaround allows users to identify missing views or reconstruction failures before leaving the site. Confidence-aware iPhone LiDAR supervision further improves PSNR from 19.874 to 20.646 dB, SSIM from 0.357 to 0.381, and depth MAE from 0.0831 to 0.0237 m. This framework establishes an efficient feedback loop for field measurements before high-accuracy offline processing.

16:45–17:00 Lightweight Visual Odometry System with a Feature-Extractable CMOS Image SensorAWARD Kazuya Yamamoto (Ritsumeikan University), Kota Yoshida (AIST), Yohei Hori (AIST), Shunsuke Okura (Ritsumeikan University)

Abstract

We propose a lightweight visual odometry system supposing a feature-extractable CMOS image sensor (FE-CIS) to address high-volume data transfer and computational overhead. The FE-CIS extracts edge features on-chip and transfers the reduced data. To reduce the processing load, the edge-intensity-based mask limits feature detection to strong edge regions. A conventional approach using full-size grayscale images achieves an absolute trajectory error (ATE) of 3.06 m. Simulation results show that the proposed system reduces data transfer volume to 12.5% and decreases the scanned pixel ratio to 22% in the feature detection stage, while limiting the ATE degradation to only 0.91 m.

[SIS]

Smart Info-media System (4)

Sat, Oct 31 16:00–17:15Rooms 201-203 (2F)

Chair: Yoshiaki Ueda (Ryukoku University)

16:00–16:15 Extending RemixIT with a Multi-Scale PatchGAN Discriminator for Unsupervised Domain Adaptation of Speech EnhancementAWARD Kyosuke Hanzawa (Saitama University), Yosuke Sugiura (Meiji University), Tetsuya Shimamura (Saitama University)

Abstract

Supervised speech enhancement models can degrade on unseen domains, while supervised adaptation requires paired clean–noisy target-domain recordings. In this paper, we propose an unsupervised adaptation framework that extends RemixIT with a Multi-Scale PatchGAN discriminator. The discriminator provides a teacher-independent perceptual signal. The framework achieves a 6.7% PESQ improvement over RemixIT and also outperforms a conventional unsupervised baseline under our evaluation protocol.

16:15–16:30 Two-Speaker Simultaneous Room Impulse Response Estimation Using Safeguarded and Phase-Modulated Music SignalsAWARD Ibuki Numabe (Saitama University), Yosuke Sugiura (Meiji University), Tetsuya Shimamura (Saitama University)

Abstract

This paper proposes a two-speaker simultaneous room impulse response (RIR) estimation method using music signals. The proposed method uses a safeguarded music signal for SP1 and a phase-modulated version of the safeguarded signal for SP2. The two RIRs are separately estimated from a mixed observation signal by frequency domain regularized least squares. Simulations under white, pink, and red background noise show that the estimation error decreases as the SNR and phase modulation scale increase.

16:30–16:45 CCFMamba: A Cross-Channel Frequency Mamba Network for Two-Microphone Speech EnhancementAWARD Kaisei Suzuki (Saitama University), Yosuke Sugiura (Meiji University), Tetsuya Shimamura (Saitama University)

Abstract

We propose CCFMamba for frequency-domain two-microphone speech enhancement. It combines causal cross-channel Mamba-based gated noise subtraction with bidirectional frequency-axis modeling. Under challenging SNR conditions of −40 to −30 dB, CCFMamba outperforms Demucs-based baselines.

16:45–17:00 Evaluating the Robustness of Video Denoising Methods under Weather-Induced DegradationAWARD Shuto Ueguri (Saitama University), Yosuke Sugiura (Meiji University), Tetsuya Shimamura (Saitama University)

Abstract

CNN-based video denoising models achieve high performance on real camera noise but often struggle with weather-included degradation. This paper investigates the adaptation of standard denoising methods to fog and snow conditions. Our analysis reveals that lightweight methods highly optimized for synthetic sensor noise face structural limitations when adapting to weather-induced noise.

17:00–17:15 SHAP Analysis of Global Visual Features for Luxury Classification in Interior ImagesAWARD Hinako Imamura (Doshisha University), Daichi Iwadare (Doshisha University), Masahiro Okuda (Doshisha University)

Abstract

This study investigates the visual factors associated with the classifier-derived luxury score in interior images. Using the logit output of a fine-tuned MobileNetV2 classifier as a continuous luxury score, we built a surrogate regression model using 17 statistically selected image features covering color, texture, and composition. An ensemble of Random Forest, XGBoost, and LightGBM achieved a 5-fold average R^2 of 0.6520. SHAP analysis revealed that Complexity Density, saturation, and depth-of-field clarity positively contribute to the luxury score, while texture directionality and uniformity contribute negatively, suggesting the luxury score arises from interacting visual factors.

[RCC]

Reliable Communication and Control

Sat, Oct 31 16:00–17:15Hall (2F)

Chair: Ryosuke Adachi (Yamaguchi University)

16:00–16:15 Vector Similarity Search Based Deep Unfolding for Generalizable Consensus ControlAWARD Tomoya Okamura (Kagawa University), Koji Ishii (Kagawa University)

Abstract

To enhance the generalization capability of data-driven optimization networks for consensus control over unseen graph realizations generated from the same network model, this paper proposes a Vector Similarity Search (VSS)-based approach. Prior work imposed centrality-based clustering constraints on edge weights, but this discretization forces a trade-off in the cluster number, since coarse clustering limits expressiveness and fine clustering causes overfitting. By leveraging VSS, we directly exploit continuous centrality values to estimate edge weights, eliminating the cluster-number trade-off.

16:15–16:30 Decentralized DQN Design for Slotted-ALOHA based Average Consensus ControlAWARD Kuninari Mizu (Kagawa University), Koji Ishii (Kagawa University)

Abstract

This paper proposes a Deep Q-Network (DQN)-aided Slotted-ALOHA for average consensus control, where the DQN decides a transmit or idle action per time slot. Since consensus control is a distributed control scheme based only on information received from neighbor agents, global network information is unobservable, and thus the DQN must be designed in an autonomous and decentralized manner. To this end, the proposed DQN relies only on local information: both the reward and input states are composed of local information. Simulation results show that the proposed method improves convergence performance compared with conventional degree-based fixed-probability Slotted-ALOHA.

16:30–16:45 TDMA Scheduling Based on Sparse ControlAWARD Naoya Yonemura (Kagawa University), Koji Ishii (Kagawa University)

Abstract

This work proposes a greedy scheduling method based on sparse control for networked control systems with limited communication resources, where only one plant can update its input at each step. The proposed method reduces total fuel consumption while keeping the entire multi-plant system stable. Simulation results demonstrated that the proposed method reduced total fuel consumption by 31.6% compared to conventional LQ sequential scheduling, while achieving full stabilization of all plants. Furthermore, a 10.0% reduction in communication updates was observed as a by-product of the sparse input structure.

16:45–17:00 Mean-Square Stability Analysis of Event-Triggered LQR Control Under Packet LossesAWARD Takumi Tamaki (Kagawa University), Koji Ishii (Kagawa University)

Abstract

This paper investigates event-triggered LQR control under packet losses. A discrete-time linear system with an ACK-based communication channel is considered. Mean-square stability (MSS) analysis is employed to derive the theoretical packet-loss tolerance. Furthermore, the relationship between the trigger threshold and packet-error rate (PER) is analyzed through an effective loss model. The analysis shows that the maximum tolerable PER decreases monotonically with the trigger threshold, which is confirmed by numerical simulations, providing a quantitative guideline for threshold design under a given communication quality.

17:00–17:15 Enhanced Resistance for Hybrid Cyber Attacks in Power System State EstimationAWARD Shuntaka Goto (Hokkaido University), Koichi Kobayashi (Hokkaido University), Yuh Yamashita (Hokkaido University)

Abstract

In power networks, enhancing the resistance to cyber attacks is important. In this paper, we propose evaluation metrics for hybrid cyber attacks that combine a False Data Injection attack with DoS attacks. We also propose an enhanced resistance method in which virtual sensors are introduced in communication networks. The proposed method is demonstrated through a numerical example on the IEEE 14-bus system.

Poster SessionSun, Nov 1 · 10:10–11:10 · Rooms 201-203 (2F)

Poster presentations are not included in the proceedings.

1
[US]Levitation force characteristics near a concave vibrating surfaceKohei Aono (Muroran Institute of Technology), Manabu Aoyagi (Muroran Institute of Technology), Hidekazu Kajiwara (Muroran Institute of Technology)
2
[MSS]Bidirectional Japanese–Okinawan Machine Translation with mT5Chuankai Dai (Institute of Science Tokyo)
3
[RCC]Event-triggered Control for Stochastic Switching Systems with Piecewise-constant Transition ProbabilitiesYuta Takahashi (Hokkaido University), Koichi Kobayashi (Hokkaido University), Yuh Yamashita (Hokkaido University)
4
[RCC]Augmented Degree-Based Design of Sparse Controller for Consensus Systems with Second-Order DynamicsKakeru Kubo (Yamaguchi University), Ryosuke Adachi (Yamaguchi University), Yuji Wakasa (Yamaguchi University)
5
[SIP]Generating Wireless Signal Using a VAE with Statistical Latent SamplingMuhammad Taqiyuddin (Kyushu Institute of Technology), Yuhei Nagao (Kyushu Institute of Technology), Masayuki Kurosaki (Kyushu Institute of Technology), Hiroshi Ochi (Kurume Institute of Technology)
6
[IT]Evaluation of Decoding Performance of a Genetic Algorithm Using a Syndrome Fitness Function in a DNA Storage EnvironmentYuta Ko (Nagaoka University of Technology), Takahiro Ota (Senshu University), Akiko Manada (Nagaoka University of Technology)
7
[IT]Symmetry-Based Leakage Bounds for a Leech-Lattice Fuzzy Extractor in PUF AuthenticationJun Asatani (Okayama University of Science)

Technical Program

Program at a Glance

Opening Fri, Oct 30 12:50–13:00

Invited Talk 1 Fri, Oct 30 13:00–14:00

Slot 1 Fri, Oct 30 14:15–15:30

Slot 2 Fri, Oct 30 15:45–16:45

Welcome Reception Fri, Oct 30 18:00–20:00

Invited Talk 2 Sat, Oct 31 09:00–10:00

Invited Talk 3 Sat, Oct 31 10:20–11:20

Slot 3 Sat, Oct 31 13:00–14:15

Slot 4 Sat, Oct 31 14:30–15:45

Slot 5 Sat, Oct 31 16:00–17:15

Invited Talk 4 Sun, Nov 1 09:00–10:00

Poster Session Sun, Nov 1 10:10–11:10

Award Ceremony Sun, Nov 1 11:30–12:00

Closing Sun, Nov 1 12:00–12:10