| 宗片 吉史 | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: *** Effective Data Allocation with 3D Stacked DRAM for LLM Acceleration on CGLA ***
abstract: *** CGLA, which combines programmability, short compilation time, and energy efficiency, has shown potential as a new dataflow computing platform capable of supporting diverse AI workloads such as LLM inference, speech recognition, and image generation. On the other hand, when executing large-scale Transformer models, the limited local memory capacity of each PE and limited bandwidth cause DMA access overhead to accumulate, resulting in a memory bottleneck. 3D-stacked DRAM is expected to serve as a technology for significantly expanding both DRAM capacity and bandwidth as semiconductor scaling approaches its limits. However, efficient memory architectures designed for TSVs and dataflow architectures remain an active area of research. In this study, we investigate an optimal data allocation method for accelerating computation using 3D-stacked DRAM, based on execution traces of Transformer-based LLM inference on CGLA. Specifically, we focus on the access frequencies of weights, activations, KV cache, and TSV bandwidth utilization. Based on these factors, frequently accessed data are preferentially placed in high-bandwidth regions. Furthermore, by introducing efficient data transfer scheduling, this study aims to reduce DMA access overhead. *** language of the presentation: *** English *** 発表題目: *** CGLA上でのLLM推論の高速化に向けた3D積層DRAMを用いた効率的データアロケーション *** 発表概要: *** プログラマビリティ、短いコンパイル時間、省電力性を兼ね備えたCGLAは、LLM推論、音声認識、画像生成といった多様なAIワークロードに対応可能な新たなデータフロー型計算基盤としての可能性を示している。一方で、大規模なTransformerモデルを実行する際には、各PEのローカルメモリ容量の制約によりDMAアクセスのオーバーヘッドが蓄積し、メモリ律速となる問題が存在する。 3D積層DRAMは、半導体微細化が限界に近づきつつある中で、DRAMの容量および帯域を大幅に拡張する技術として期待されている。しかし、TSVやデータフローアーキテクチャを前提とした効率的なメモリアーキテクチャについては、依然として研究途上である。 本研究では、CGLA上におけるTransformer LLM推論の実行トレースを基に、3D積層DRAMを用いた計算処理の高速化に向けて、最適なデータアロケーション手法を検討する。具体的には、重み、Activation、KV cacheへのアクセス頻度およびTSV帯域の利用効率に着目し、頻繁にアクセスされるデータを広帯域領域に優先的に配置する。さらに、効率的なデータ転送スケジューリングを導入することで、DMAアクセスオーバーヘッドの削減を目指す。 *** | ||||||||
| BUI NGOC THANH BINH | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: *** Protocol-Aware MQTT Security and Anomaly Mitigation in a P4 Data Plane for Edge IoT Systems ***
abstract: *** MQTT is the dominant lightweight publish-subscribe protocol for IoT deployments yet edge security remains inadequate. Cloud-based intrusion detection systems add latency unsuitable for real-time control while CPU-bound firewalls and generic SDN controllers lack MQTT awareness to enforce session validation topic-based authorization and behavioral anomaly detection. We propose a P4-based data-plane enforcement scheme for protocol-aware MQTT security and anomaly detection at the network edge. The design combines parser-safe MQTT header extraction with session-order validation byte-level topic-prefix authorization with per-client rate limiting and soft-cap enforcement and lightweight anomaly detection based on KeepAlive and Remaining Length screening with clone-to-CPU diagnostics. The scheme leverages BMv2 stateful primitives including registers meters and direct counters to enable runtime policy adaptation with minimal per-packet latency. Experiments on a Mininet-BMv2 testbed demonstrate high policy enforcement accuracy of 99.8% within 95% CI strong anomaly detection sensitivity of 98% true-positive rate and high packet delivery of greater than 99.9% for 100 to 5 kpps 99.8% at 10 kpps and 99.6% at 16 kpps with sub-millisecond per-packet latency. These results show that protocol-aware MQTT filtering can be efficiently realized in the programmable data plane providing a practical foundation for edge IoT security. Future work will validate the design on production P4 hardware and integrate machine learning-based threshold adaptation. *** language of the presentation: *** English *** | ||||||||
| LE TRAN XUAN HIEU | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: *** Second-Order Regularized Optimization for Sparse-View CT Reconstruction ***
abstract: *** Sparse-view computed tomography (CT) reduces radiation exposure by acquiring fewer projection views, but it also makes image reconstruction highly ill-posed and prone to severe streak artifacts. In this work, we propose CG-GLORE, a compact second-order deep unrolling framework for sparse-view CT reconstruction. Each reconstruction stage is formulated as a Newton-Raphson-inspired update, where the curvature of the data-fidelity term is efficiently exploited through a Conjugate Gradient solver without explicitly constructing or inverting the Hessian. To provide effective learned regularization, we introduce a Global-Local Regularization Network (GLORE), which combines convolutional local feature extraction with long-range dependency modeling based on sparse patchification and Nyström attention. This design enables the model to preserve fine anatomical details while maintaining global structural consistency. Experiments on AAPM and DeepLesion datasets under multiple sparse-view and noisy acquisition settings demonstrate that CG-GLORE achieves strong quantitative performance, stable convergence, reduced residual noise, and improved visual fidelity compared with representative reconstruction methods. Ablation studies further confirm that the CG-based second-order update and the global-local regularization module provide complementary benefits. Overall, CG-GLORE offers an effective balance between model-based optimization and learned regularization for robust sparse-view CT reconstruction. *** language of the presentation: *** English *** | ||||||||
| LI LINGWEI | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: A Fine-Grained Token Compression Mechanism for Efficient EEG Model Inferencing
abstract: Recent adaptations of Transformers for EEG tasks, such as sleep staging and seizure detection, have shown great promise. However, the quadratic computational complexity of the self-attention mechanism poses a significant challenge: processing long-duration, multi-channel EEG signals generates an excessive number of tokens, making real-time inference computationally prohibitive. To alleviate this, we propose a fine-grained token compression framework tailored for the unique properties of EEG signals. We argue that existing single-strategy methods are sub-optimal for EEG data due to its low signal-to-noise ratio: redundancy-based pruning may lead to the loss of sparse but critical neurological features, while importance-based merging risks burying meaningful information under dominant noise. To bridge this gap, our framework categorizes tokens into four distinct quadrants by jointly evaluating their importance and redundancy, subsequently applying specialized processing strategies to each category. This approach ensures a precise balance between computational efficiency and the preservation of essential diagnostic information. language of the presentation: English | ||||||||
| 大越 涼太 | M, 1回目発表 | サイバーレジリエンス構成学 | 門林 雄基 | 林 優一 | 妙中 雄三 | |||
|
title: An Investigation of Living off the Agent Attacks and Defense Mechanisms in AI Agent Systems
abstract: Large Language Model (LLM)-based AI agents have recently been integrated with external tools, memory systems, and web services to automate complex tasks. However, these capabilities introduce new security risks. In particular, Living off the Agent (LoTA) attacks exploit legitimate agent functions and tool interactions to achieve malicious objectives without relying on traditional software vulnerabilities. Existing guardrail mechanisms primarily focus on filtering user inputs and model outputs, which may be insufficient for detecting risks emerging from multi-step agent workflows. This study investigates the characteristics of LoTA attacks in AI agent environments and analyzes the limitations of current defense mechanisms. Through the analysis of representative attack scenarios, we aim to identify security challenges specific to AI agent workflows and discuss requirements for effective countermeasures. language of the presentation: *** English or Japanese (choose one) *** English 発表題目: AI AgentにおけるLiving off the Agent（LoTA）攻撃とその対策に関する検討 発表概要: 近年，大規模言語モデル（LLM）を活用したAI Agentは，外部ツールやWebサービスと連携することで複雑なタスクを自律的に実行できるようになっている。一方で，これらの機能を悪用した新たな攻撃手法が報告されており，その一つとしてLiving off the Agent（LoTA）攻撃が注目されている。LoTA攻撃は，従来のソフトウェア脆弱性を利用するのではなく，AI Agentが本来備える機能や権限を悪用することで攻撃を実現する特徴を持つ。 本研究では，AI Agent環境におけるLoTA攻撃の特徴を整理するとともに，既存のガードレール機構や防御手法の課題について分析する。さらに，代表的な攻撃シナリオを通じて，AI Agentのワークフローにおいて発生し得るセキュリティ上の問題点を明らかにし，今後必要となる対策について検討する。 | ||||||||
| DU CAN | M, 1回目発表 | 生体画像知能 | 大竹 義人 | 金谷 重彦 | Gu Yi | Soufi Mazen | ||
|
title: *** AI-Driven Registration-Based Multi-Task Model for 3D Muscle Fiber Orientation, Volume, and Intramuscular Fat Prediction from Routine CT/MRI with DTI ***
abstract: *** Sarcopenia and muscle degradation are critical factors impacting the independence of the aging population. While clinical CT/MRI can effectively assess muscle volume and fatty infiltration, they fail to directly predict muscle fiber orientation, which is fundamental to muscle strength and function. Although Diffusion Tensor Imaging (DTI) enables 3D visualization of muscle fibers, its clinical application is constrained by long scan times. This study aims to develop an AI-driven model that predicts 3D fiber orientation, muscle volume, and fat fraction from routine CT/MRI scans. We established a paired dataset of DTI and clinical CT/MRI scans, and developed:(1) automated segmentation of lower limb muscles from MRI to quantify volume and fat fraction; (2) construction of a registration-based musculoskeletal model integrated with DTI-derived diffusion directions to generate ground-truth 3D fiber orientation maps; and (3) a multi-task learning framework to map MRI data to these architectural metrics. This approach transforms routine clinical imaging into comprehensive functional assessments, offering a non-invasive, efficient tool for the objective evaluation of muscle quality and clinical decision-making. *** language of the presentation: *** English *** | ||||||||
| AFNAN SUKRI ANWAR ABDULLAH | M, 1回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | Sakriani Sakti | 若宮 翔子 | 西山 智弘 | ||
|
title: *** Exploring Novel Drug Research Area using Large Language Models Based on Research Trends in Biomedical Literature ***
abstract: *** The rapid expansion of biomedical literature makes manual identification of novel drug-disease relationships increasingly difficult. Existing approaches have leveraged LLMs to mine abstracts or construct knowledge graphs for drug repurposing. There are two key limitations: finite context windows for capturing macro-level research trends, and single-pass black-box pipelines make it difficult to verify outputs. This paper proposes a pipeline for discovering new drug targets by combining disease and drug research trends using Large Language Models (LLMs). Our method extracts PICO components from PubMed abstracts, normalizing the Population and Intervention Component to ICD and ATC codes, respectively. A temporal frequency delta matrix is constructed to capture publication count shifts across 2013 to 2022, then used to discover novel drug areas. Compared with the abstract-based baseline, our approach showed qualitative signs of generating combinations that were more closely aligned with observed research trends and, in some cases, more clinically plausible. These findings suggest the potential usefulness of structured trend information for LLM-based exploration, although the differences between the two methods were limited and the results remain preliminary. Future work will focus on validating the consistency and reliability of these candidates. *** language of the presentation: *** English *** | ||||||||
| DAI NING | M, 1回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | 安本 慶一 | 若宮 翔子 | 久田 祥平 | ||
|
title: *** Classification of Women’s Health-Related Posts and Structuring of Symptom Expressions Using a Large Language Model ***
abstract: *** This study used texts posted on “Hatsugen Komachi,” an online discussion board for women, as its target data. Hatsugen Komachi is an online forum where users exchange advice and opinions on everyday life topics, including daily living, work, family, romantic relationships, pregnancy and childbirth, and health. The posts describe users’ own experiences and concerns in natural language, and may include physical discomfort and anxiety before visiting a medical institution, as well as health-related difficulties that are difficult to discuss with others. In this study, each post was treated as the unit of analysis. The input text consisted of the main topic text and, when necessary, response texts. The target posts included descriptions of the poster’s own physical condition, experiences of medical consultation, consultation with medical institutions, anxiety about diseases and treatment, concerns related to pregnancy, childbirth, and infertility, and physical or mental distress associated with family relationships, childcare, or caregiving. *** language of the presentation: *** Japanese *** | ||||||||
| BAH SADIO | M, 1回目発表 | ソフトウェア工学 | 松本 健一 | 安本 慶一 | 嶋利 一真 | Fan Youmei | ||
|
title: Quantization Effects on LLM-Based Coding Agents for Automated Bug Detection, Repair, and Testing
abstract: Large Language Model (LLM)-based coding agents have recently demonstrated strong capabilities in software engineering tasks such as bug detection, repair, and automated testing. However, deploying these models often requires substantial computational and memory resources. Quantization is a widely used technique that reduces model size and computational cost by lowering numerical precision, enabling more efficient deployment. Despite these advantages, the impact of quantization on the debugging performance of LLM-based coding agents remains insufficiently understood. This research investigates how different quantization levels (FP16, INT8, and INT4) affect the performance and efficiency of LLM-based coding agents using the BugsInPy benchmark. The study evaluates the impact of quantization across different debugging stages and bug types, while also examining efficiency-related factors such as runtime and memory usage. Through this analysis, this research aims to provide a better understanding of the trade-offs between debugging effectiveness and computational efficiency under different quantization levels. language of the presentation: English | ||||||||
| LAOHAKANNIYOM PAKORN | M, 1回目発表 | ソフトウェア設計学 | 飯田 元 | 松本 健一 | 柏 祐太郎 | Reid Brittany | ||
|
title: Your Latest is Not My Latest! An Empirical Study of Latest Tags in Dockerfiles
abstract: The latest tag in Docker offers developers a convenient shorthand for referencing base images without track- ing explicit version numbers. However, because this tag is a mutable, shifting reference, Dockerfiles that depend on it are silently exposed to breaking changes whenever the upstream image is updated. This creates a reproducibility gap that directly undermines one of containerization core promises. Despite widespread recognition of version pinning as best practice, the prevalence and consequences of latest usage in real-world projects remain poorly understood. This paper presents an empirical study of latest tag usage across 21,481 Dockerfiles drawn from 9,179 open-source GitHub repositories. We find that 14.2% of all Dockerfiles and 24.0% of repositories depend on latest for actively maintained, multi- release projects. When developers do revise their tagging strategy, they are 64% likely to migrate away from latest toward a pinned version than in the opposite direction, suggesting awareness of the risks or necessity after failures are encountered. Additionally, through dynamic build analysis, we measure how quickly Dockerfile buildability degrades as base images evolve language of the presentation: English | ||||||||
| DAMARHAFNI RAHMANNABEL NADIM PRAMONO | M, 1回目発表 | ヒューマンAIインタラクション | Sakriani Sakti | 渡辺 太郎 | 大内 啓樹 | Faisal Mehmood | Bagus Tris Atmaja | |
|
title: *** Multi-Stream Artifact Detection for Audio Deepfake Detection via Temporal, Spectral, and Phase Cues ***
abstract: *** Audio deepfakes produced by modern text-to-speech and voice-conversion systems pose growing threats to digital security and media trust. While existing detection methods often rely on a single feature domain, they frequently fail to generalize across diverse synthesis algorithms and acoustic conditions. We propose the Artifact Detection Model (ADM), a lightweight multi-stream neural classifier that decomposes deepfake detection into three complementary artifact domains: temporal, spectral, and phase. Each domain is processed by a dedicated residual MLP sub-module operating on mean-pooled, utterance-level features derived from self-supervised learning embeddings, mel-filterbank energies, and STFT phase features. We systematically investigate multiple stream-fusion strategies across four diverse benchmarks: ASVspoof 2019 LA, ASVspoof 2021 LA, ASVspoof 2021 DF, and the In-the-Wild dataset. While the temporal branch serves as the strongest individual stream, a three-branch fusion utilizing element-wise max aggregation achieves superior robustness, yielding highly competitive Equal Error Rates (EERs) across both controlled benchmarks and unconstrained real-world datasets. These results demonstrate that fusing low-level hand-crafted phase cues with high-level temporal embeddings successfully mitigates performance degradation across unseen acoustic conditions. *** language of the presentation: *** English *** | ||||||||
| RUDY ONG | M, 1回目発表 | ヒューマンAIインタラクション | Sakriani Sakti | 渡辺 太郎 | 大内 啓樹 | Faisal Mehmood | Bagus Tris Atmaja | |
|
title: *** ASR-Based Detection of Vowel Devoicing in Japanese Speech Corpora ***
abstract: *** The Japanese-Language Proficiency Test (JLPT) is widely recognized among foreign workers in Japan. However, it limited to only evaluates reading and listening skills and does not assess spoken proficiency. As effective workplace communication relies heavily on accurate speech production and pronunciation, additional tools are needed to support learners' verbal language development. Computer-Assisted Pronunciation Training (CAPT) systems leverage Automatic Speech Recognition (ASR) to provide pronunciation support for second-language (L2) learners. A key requirement for effective pronunciation feedback is the accurate detection of specific pronunciation phenomena. Despite its importance in Japanese speech, vowel devoicing has received limited attention in CAPT research.Therefore, this study proposes a method for detecting Japanese vowel devoicing from input speech by representing devoiced vowels as explicit IPA tokens within lexical representations. To evaluate detection performance, we introduce two metrics: Devoicing Event Occurrence (DEO) and Context-Conditioned Devoicing Accuracy (CCDA). Experimental results demonstrate strong performance on both metrics. These findings suggest that explicit detection of vowel devoicing can serve as a foundation for incorporating pronunciation feedback on vowel devoicing into Japanese CAPT systems. *** language of the presentation: *** English *** | ||||||||
| AN MINHYOUNG | M, 1回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 荒牧 英治 | 諏訪 博彦 | 佐々木 航 | ||
|
title: Establishing a Grounded Context Inference Method for Large Language Models
abstract: Advances in large language models (LLMs) have made it possible to infer context from sensor data in natural language. However, such inference is typically evaluated by its agreement with a ground truth described by the users themselves (user annotation). Let the real context R denote the set of information that can be regarded as true in a given situation. In the real world, R is multifaceted, and the annotation S cannot fully cover R due to cognitive and temporal constraints (omission). Meanwhile, the model output M contains outputs not included in R (hallucination). Because the omission in S and the hallucination in M occur simultaneously, the outputs outside the annotation (M−S) contain a mixture of outputs that are included in R (grounded inferences) and hallucinations, and the two cannot be distinguished. This study focuses on the set (M−S)∩R, which can compensate for both limitations, and defines it as the grounding region. Starting from this grounding region, the research proceeds in two directions. First, we identify the grounding region and clarify its structure. Second, without increasing the number of user responses or the volume of model output, we construct a bidirectional framework that uses the grounding region as a foothold to bring both S and M closer to the real context R. Through this, we aim to enable context inference that reaches even the aspects users have not verbalized, contributing to natural interaction between people and systems. language of the presentation: Japanese 発表題目: 接地領域に基づく大規模言語モデルのコンテキスト推論手法の確立 発表概要: 大規模言語モデル（LLM）の発展により，センサデータからコンテキストを自然言語で推論することが可能となった．しかしその評価は，ユーザ自身が記述する正解（ユーザアノテーション）との一致に依存している．ここで，実コンテキストR を対象状況において真とみなせる情報の集合とすると，実世界の R は多面的であり，アノテーション S は認知的・時間的制約から R を網羅できない（欠落）．一方，モデル出力 M は R に含まれない出力（ハルシネーション）を伴う．S の欠落と Mのハルシネーションが同時に存在するため，アノテーション外の出力 (M−S) には，R に含まれる出力（接地した推論）とハルシネーションが混在し，両者を区別できない． 本研究は，この両者の限界を補い得る集合 (M−S)∩R に着目し，これを「接地領域」と定義する．接地領域を起点として，本研究は二つの方向から研究を進める．第一に，接地領域を特定し，その構造を明らかにする．第二に，ユーザの回答数とモデルの出力量を増やすことなく，接地領域を足がかりにS と M の双方を実コンテキストR へ接近させる双方向フレームワークを構築する．これにより，本人が言語化していない側面にまで踏み込んだコンテキスト推論を実現し，人とシステムの自然なインタラクションに寄与する． | ||||||||
| ADNAN IMRANUL ISLAM | M, 1回目発表 | 光メディアインタフェース | 向川 康博 | 安本 慶一 | 藤村 友貴 | 北野 和哉 | ||
|
title: Just Photos: 3D reconstruction from Sparse Unposed Images
abstract: This research will investigate a method for reconstructing high-quality 3D scenes from a small set of sparse, unordered photographs without requiring camera parameters or traditional structure-from-motion pipelines. The proposed approach will combine monocular depth estimation and pose prediction models to automatically recover scene geometry and camera viewpoints from images alone. A symmetric depth consistency loss will be introduced to align depth information across multiple views and reduce geometric artifacts such as duplicated or misaligned structures caused by independently estimated depth maps. In addition, a depth-guided test-time refinement strategy will be explored to further improve scene geometry using newly observed images. The effectiveness of the proposed method will be evaluated on indoor and outdoor datasets using both reconstruction quality and novel-view synthesis performance. English | ||||||||
| HASAN S M MEHEDI | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎 | 荒牧 英治 | 上垣外 英剛 | 坂井 優介 | ||
|
title: Analysis on Generalization Ability of Translation Evaluation Models through Layer-wise Cross-lingual Error Representation Analysis
abstract: Automatic evaluation models are widely used to assess machine translation quality, but their internal behavior across languages and error conditions remains unclear. This study analyzes the generalization ability of translation evaluation models through layer-wise cross-lingual error representation analysis. We compare XLM-RoBERTa, a pretrained multilingual encoder, with xCOMET, a fine-tuned translation evaluation model, using XQ-MEval as a controlled multilingual dataset with injected MQM-defined errors and error-span information. We extract sentence-level and span-level representations using masked average pooling and mainly apply PCA to examine language-pair geometry, layer-wise changes, quality/error-count alignment, score alignment, and error-span organization. Through this analysis, we aim to understand how MT-evaluation fine-tuning reshapes internal representations and whether error-sensitive structures are shared across language pairs. language of the presentation: English | ||||||||
| LIU TIANYU | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: Accuracy and Reliability Evaluation of Low-Bit Spiking Quantized Large Language Models under Hardware Faults
abstract: Low-bit quantization is an important technique for reducing the memory footprint and inference cost of large language models. In my previous work, I studied teacher-guided saliency refinement for low-bit spiking quantized LLMs, where full-precision layer outputs are used as teacher signals to refine channel-wise saliency estimation and improve high-precision channel allocation. Experimental results on LLaMA-2-7B under the W4A4 setting showed that this method can improve perplexity on WikiText2 and C4, as well as average accuracy on several zero-shot reasoning tasks. Based on this work, my current research focuses on the reliability of low-bit spiking quantized LLMs under hardware transient faults. Since spiking-inspired quantization represents activations through temporal or branch-based encoding, it may change the error propagation behavior compared with conventional quantized LLMs. I plan to construct a fault-injection evaluation framework and analyze how different layers, modules, fault probabilities, and fault types affect model outputs and task accuracy. The goal of this research is to clarify the relationship among low-bit quantization, saliency-aware precision allocation, and hardware fault tolerance, and to explore whether spiking-based LLM quantization can provide a better accuracy, efficiency, and reliability trade-off. language of the presentation: English | ||||||||
| PHAM ANH KIET | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: *** HMSA: High-Performance Multi-Mode SPHINCS+ Accelerator With a 16-Core SHA-256 Engine on FPGA SoC ***
abstract: *** SPHINCS+ is a stateless hash-based digital signature scheme for post-quantum security, but its practical use on embedded platforms is limited by the large amount of repeated hashing required across key generation, signing, and verification. This paper presents HMSA, a high-performance multi-mode SPHINCS+ accelerator for FPGA SoCs. HMSA supports the complete SPHINCS+ flow, including key generation, signature generation, and signature verification, for the six SHA-256 simple parameter sets: 128f, 128s, 192f, 192s, 256f, and 256s. The proposed architecture uses a shared 16-core SHA-256 engine with a unified control and data path, allowing independent hash tasks from WOTS+, FORS, and XMSS to be processed in parallel while keeping the design scalable across parameter modes. We implemented HMSA on a Xilinx ZCU102 FPGA SoC and compared it with single-core, 4-core, and 8-core architectures. Experimental results show that HMSA reduces signing cycles by 7.1x-13.0x over the single-core baseline, 3.1x-3.6x over the 4-core design, and 1.8x-2.2x over the 8-core design. Across the supported modes, key generation and verification are also accelerated by up to 14.7x and 7.6x, respectively. These results demonstrate that HMSA provides an efficient and scalable hardware platform for full SPHINCS+ acceleration on FPGA SoCs. *** language of the presentation: *** English *** | ||||||||
| SHENG GUAN | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: LoRA-based LLM Reliable Deployment on Memristor Crossbars
abstract: Memristor crossbars provide a promising platform for energy-efficient deployment of large language models (LLMs) by performing matrix-vector multiplication directly in memory. However, analog devices has limited precision and analog peripheral circuits such as analog-to-digital converters (ADCs) may introduce additional noises. And the conductance may variant when voltage press on it for computation, which may cause the outputs of LLMs totally changed. Low Rank Adaptation (LoRA) was used for LLM finetune and recently are used for enhance the reliability of analog devices by adding noise to W and training LoRA to recover. My research is aiming to measure the performance after the adaptive training and exploring the possibility of proposing better software-hardware co-design solutions. language of the presentation: English | ||||||||
| YUCHANG HUANG | M, 1回目発表 | コンピューティング・アーキテクチャ | 中島 康彦 | 林 優一 | 張 任遠 | KAN Yirong | PHAM HOAI LUAN | Le Vu Trung Duong |
|
title: A Topology-Aware Spiking Transformer for Sparse sEMG-Based Gesture Recognition
abstract: Surface electromyography-based hand gesture recognition plays an important role in prosthetic control, rehabilitation assistance, and human–computer interaction. Although CNNs, TCNs, and Transformers have improved recognition accuracy, their high computational cost and energy consumption remain challenging for wearable and embedded applications. This research explores a topology-aware Spiking Transformer framework for sparse sEMG-based gesture recognition. The proposed research focuses on three aspects. First, a spike encoder with explicit rules is investigated to stabilize spike activities at both the channel level and the local patch-token level. Second, topology-aware analysis is introduced to address the irregular electrode layout of sparse sEMG, where meaningful tokens are not naturally defined. By constructing a virtual electrode topology, local spike tokens can better preserve spatial relationships among electrodes. Third, lightweight spike attention mechanisms are explored to reduce redundant token interactions and improve computational efficiency. language of the presentation: English | ||||||||
| GOTMARE YUGANT RAJKUMAR | M, 1回目発表 | 生体画像知能 | 大竹 義人 | 冨谷茂隆（物質） | Gu Yi | Soufi Mazen | ||
|
title: Seeing Joints in Motion: Articulated 2D-3D Fluoroscopy and CT Registration for Knee Kinematics Estimation with Biomechanical Constraints
abstract: Understanding how the knee moves during activities such as standing, bending, and squatting is important for studying joint function and supporting medical research. One common approach is to align 2D fluoroscopic X-ray images with 3D models created from CT scans; however, achieving accurate and stable alignment throughout a motion sequence remains a challenging task. In this project, we present an articulated 2D–3D fluoroscopy-to-CT registration method for estimating knee joint motion. The method tracks the movement of the femur, tibia, and patella throughout a motion sequence. To improve registration stability and maintain anatomically realistic motion, biomechanical constraints are incorporated into the optimization process. A sequential tracking strategy is also used, allowing information from previous frames to guide the registration of subsequent frames. The framework is currently under development, and preliminary experiments indicate that incorporating biomechanical constraints can help produce more consistent joint motion estimates while reducing anatomically unrealistic solutions. Ongoing work focuses on further improving registration robustness and evaluating the method on dynamic knee motion datasets. language of the presentation: English | ||||||||
| JIN YEDONG | M, 1回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | 渡辺 太郎 | 若宮 翔子 | PENG SHAOWEN | ||
|
title: Why Alignment Falls Short in LLM-Enhanced Recommendation: A Spectral
Discrepancy View
abstract: Recent advances in large language models (LLMs) have improved semantic-enhanced recommender systems by introducing richer language representations that can be aligned with collaborative signals. Despite promising empirical results, the mechanism by which semantic representations improve recommendation performance remains poorly understood. In this work, we revisit LLM-enhanced recommendation from a spectral perspective and uncover a fundamental discrepancy between collaborative and semantic signals. Specifically, while collaborative representations are typically dominated by smooth low-frequency components due to user-item interaction homophily, semantic embeddings contain substantial informative non-dominant components that are important for recommendation performance. In contrast, we empirically show that commonly adopted alignment-based objectives tend to emphasize low-frequency semantic and collaborative components while suppressing non-dominant (high-frequency) semantic ones throughout training, exhibiting a tendency similar to that of interaction-only CF methods. Motivated by these observations, we introduce a decoupling paradigm that preserves semantic and collaborative signals in their native spaces instead of forcing them into a shared embedding space. The consistent improvement of the decoupled paradigm over alignment-based counterparts further supports its effectiveness. Based on this idea, we propose UniSpecRec (\textbf{Uni}fying \textbf{Spec}tral Signals for \textbf{Rec}ommendation), a simplified and parameter-free framework that integrates collaborative and semantic signals in the spectral domain. Extensive experiments on multiple benchmark datasets and LLM encoders demonstrate the effectiveness, efficiency, and generalizability of the proposed approach. language of the presentation: English | ||||||||
| MARPENA ADRIANNA | M, 1回目発表 | ソフトウェア工学 | 松本 健一 | 安本 慶一 | 嶋利 一真 | Fan Youmei | ||
|
title: Evaluating AI-Generated vs Human-Written SECURITY.md Files in Open-Source Repositories
abstract: Security documentation is essential in open-source software, guiding responsible vulnerability disclosure. With the growing use of large language models (LLMs) in software engineering, AI is increasingly used to generate documentation, including SECURITY.md files. However, it remains unclear whether AI-generated files are as complete and reliable as those written by humans. This study evaluates how closely AI-generated SECURITY.md files match human-written versions in terms of completeness, correctness, and reliability. It addresses two research questions: (RQ1) identifying differences between AI- and human-written files, and (RQ2) assessing AI performance in component coverage, missing information, and hallucinated content. The methodology uses an experimental design with 211 PyPI repositories. Original SECURITY.md files are removed, and AI models regenerate them using full repository context. Generated and original files are then compared using metrics such as component coverage, gap rate, hallucination rate, and semantic similarity (BERTScore). The research is currently in the experimental design phase, with dataset selection and workflow defined. Next steps include executing the experiment and analyzing results to answer the research questions. language of the presentation: English | ||||||||
| SONPEE SILA | M, 1回目発表 | ソフトウェア設計学 | 飯田 元 | 松本 健一 | 柏 祐太郎 | Reid Brittany | ||
|
title: An Empirical study on Performance-related Self-Admitted Technical Debt
abstract: Developers frequently leave comments in their code admitting known problems for example, "this is inefficient, I'll fix it later." Such comments are known as self-admitted technical debt (SATD). However, current large-scale studies classify SATD into broad categories such as design, defect, and requirement debt, while performance is never treated as a category of its own. This empirical study is the first to isolate performance-related SATD (perfSATD) from labeled SATD datasets and categorise it, examining how prevalent it is, how accurately it can be detected automatically, and what types of performance debt exist. language of the presentation: English | ||||||||
| XIA SISI | M, 1回目発表 | ソフトウェア設計学 | 飯田 元 | 松本 健一 | 柏 祐太郎 | Reid Brittany | ||
|
title: Do Agent Context Files Provide Security Guardrails? An Empirical Study of Agentic AI Software Repositories
abstract: Agentic AI coding tools increasingly rely on repository-level context files to guide agent behavior during software development. These files often contain project-specific instructions, coding conventions, and operational constraints that influence how AI agents generate and modify code. As AI coding agents gain greater autonomy, security-related guidance within context files may play an important role in shaping secure development practices. However, little is currently known about the extent to which these artifacts contain explicit security guardrails. In this paper, we present an empirical study of security guardrails in agent context files collected from open-source software repositories. Using a large-scale dataset of agentic AI coding tool configurations, we investigate whether and how security-related constraints are specified in context files. We develop a classification framework based on security best practices and OWASP risk categories to identify and characterize security guardrails embedded in these artifacts. Our study aims to provide insights into the prevalence and characteristics of security guidance in agent context files and to establish a foundation for future research on secure agentic software engineering. language of the presentation:Japanese Agent Context Filesにおけるセキュリティガードレールの実態分析：Agentic AIソフトウェアリポジトリを対象とした実証研究 近年、Claude CodeやCodex、CursorなどのAgentic AIコーディングツールが急速に普及している。これらのツールでは、AIエージェントの振る舞いを制御するために、リポジトリレベルのContext Files（例：AGENTS.md、CLAUDE.md、copilot-instructions.md）が利用されている。これらのファイルには、コーディング規約や開発プロセス、プロジェクト固有の制約などが記述されており、AIエージェントによるコード生成や修正に影響を与える。 AIエージェントの自律性が高まるにつれ、Context Filesに記載されたセキュリティ関連の指示は、エージェントの不適切な行動を抑制するための「セキュリティガードレール」として機能する可能性がある。しかし、既存研究では、Context Filesにどの程度セキュリティガードレールが含まれているのか、またどのようなセキュリティリスクに対応しているのかについて十分に明らかになっていない。 本研究では、オープンソースソフトウェアリポジトリから収集されたAgentic AIツール設定データセットを用いて、Context Filesに含まれるセキュリティガードレールの実態を大規模に調査する。具体的には、OWASP Top 10 for LLM Applicationsおよび既存のセキュリティガイドラインを基に分類基準を構築し、Context Filesに記述されたセキュリティ関連制約を識別・分析する。これにより、Agent Context Filesにおけるセキュリティガードレールの普及状況と特徴を明らかにし、安全なAgentic Software Engineeringに関する今後の研究基盤を提供することを目指す。 | ||||||||
| SULTANA NAYMA | M, 1回目発表 | ディペンダブルシステム学 | 井上 美智子 | 門林 雄基 | 笹田 大翔 | |||
|
title: Can Federal Learning be Trusted? A Study of Backdoor Vulnerebilities and Defences
abstract: Federated Learning (FL) enables collaborative model training without sharing raw data, making it suitable for privacy-sensitive applications such as healthcare and research. However, FL remains vulnerable to backdoor attacks, where malicious clients inject poisoned updates to compromise the global model. Existing defense mechanisms, including robust aggregation, anomaly detection, and differential privacy, often rely on restrictive assumptions such as IID data, clean validation datasets, or fixed attack patterns. As a result, they may fail against adaptive attackers and highly heterogeneous environments.This research investigates the limitations of current FL backdoor defenses and proposes an adaptive client-side defense framework. The approach focuses on detecting backdoor triggers, analyzing client update behavior, and dynamically selecting appropriate defense strategies. By integrating detection, analysis, and adaptation, the proposed framework aims to improve robustness against evolving backdoor attacks while maintaining model accuracy and privacy in non-IID federated learning environments. language of the presentation: English | ||||||||
| SHEN XIULIN | M, 1回目発表 | ヒューマンAIインタラクション | Sakriani Sakti | 渡辺 太郎 | 大内 啓樹 | Faisal Mehmood | Bagus Tris Atmaja | |
|
title: Situation-aware Embodied Conversational Agent for Disaster Guidance
abstract: This research aims to design an Embodied Conversational Agent (ECA) for VR-based disaster guidance and investigate its effectiveness in a simplified evacuation scenario. The proposed system uses SimWorld (Based on Unreal Engine) to build a small scale evacuation environment, where the ECA observes user's movement, route deviation, hesitation, and voice input to estimate basic user states. Based on these states, the agent provides adaptive route guidance, warning, reassurance, and simple explanations. The study explores how user state recognition and adaptive guidance can support disaster training, and evaluates the system through both task performance and user experience. language of the presentation: English | ||||||||
| LIU YUEXIAO | M, 1回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 岡田 実 | 諏訪 博彦 | 松井 智一 | ||
|
Title:Evaluating Self-Supervised Representation Learning for Non-Visual Activity Recognition in Real-World Smart Homes
abstract:s study evaluates non-visual activity recognition using real-world smart-home sensor data. Raw asynchronous sensor events are converted into fixed-interval wide tables and sliding-window sequences, and activity labels are assigned from manually recorded ON/OFF activity intervals. We compare BYOL+KMeans, BYOL linear probing, supervised learning from scratch, BYOL fine-tuning, and BYOL+CE under limited event-level labels. We also examine sensor-only, time-only, and sensor-plus-time settings. The results show that BYOL+KMeans is insufficient for semantic activity recovery, while linear probing indicates that BYOL embeddings contain useful activity information. This study clarifies both the utility and limitations of self-supervised learning in real smart-home HAR. language of the presentation: Japanese 発表題目: *** この部分を発表題目に *** 発表概要: *** この部分を発表概要に *** | ||||||||
| ZHANG YUANZHUO | M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦 | 松本 健一 | MD.Altaf-Ul-Amin | Ahmad Kamal Nasution | ||
|
title: Dynamics and Modulation of Slow-Wave Sleep in a Whole-Brain Mode
abstract: Slow-wave sleep (SWS) is an important physiological brain state characterized by large-scale synchronized neural oscillations and is closely associated with memory consolidation, information processing, and neural plasticity. However, many questions remain regarding how slow-wave activity emerges from local neuronal dynamics to form synchronized activity across the whole brain, and how external stimulation may influence these dynamics. The aim of this research is to develop a whole-brain computational model for investigating the dynamical mechanisms of slow-wave sleep and its modulation. The study is based on the large-scale neural model proposed by Goldman et al., which combines biologically inspired neural population models with the human structural connectome to simulate brain activity under different states. At the current stage, the simulation environment has been established, and the model structure as well as key parameters have been analyzed. Wake-like and sleep-like dynamical states have been successfully generated, including the characteristic Up-Down state transitions observed during slow-wave sleep. Future work will focus on the analysis of slow-wave characteristics, stimulus-response dynamics, and the development of modulation strategies. Ultimately, this research aims to investigate the mechanisms through which external stimulation influences slow-wave activity and to provide a foundation for future studies on closed-loop neuromodulation. language of the presentation: English | ||||||||
| LI CHUYANG | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎 | Sakriani Sakti | 上垣外 英剛 | 坂井 優介 | ||
|
title: Identifying Neurons that Control Span-Level Errors
abstract: Span-level error estimation models used in machine translation estimate translation quality by identifying errors at the span level. These models are often applied to unseen language pairs and domains at inference time, but there has been limited investigation into whether they generalize appropriately to such settings, or whether such generalization is even possible. In this study, we analyze these models at the neuron level and reveal the challenges involved in identifying span-level errors. language of the presentation: English | ||||||||