| 鎌田 亘佑 | M, 2回目発表 | 大規模システム管理 | 笠原 正治, | 林 優一, | 原 崇徳, | Christopher Wiraatmaja |
|
title: Efficient Zero-Knowledge Proofs of Gradient Descent for Quantized Neural Networks
abstract: The growing use of sensitive data in machine learning has increased the need to verify model training without revealing the training data or model parameters. Zero-knowledge proofs provide a promising approach to verifying such computations while preserving privacy. Existing works combine different proof techniques for linear and nonlinear computations, resulting in substantial prover overhead and complex composition. In this work, we present a more efficient proof architecture for gradient-descent computations. The proposed design uses LUT techniques for nonlinear computations and processes the remaining linear computations in a streaming fashion. We describe the protocol design, its integration with forward and backward propagation, and its implementation for quantized neural networks. language of the presentation: Japanese 発表題目: 量子化ニューラルネットワーク学習の勾配降下における高速ゼロ知識証明 発表概要: 機械学習における機密データの利用が拡大するにつれて、学習データやパラメータを秘匿しつつ、モデル学習の正しさを検証する必要性が高まる。ゼロ知識証明は、プライバシーを保護しながら、このような計算の正しさを検証するための有望な手段である。既存のプロトコルでは、線形計算と非線形計算に異なる証明技術を組み合わせており、その結果、証明者の計算負荷が増大するとともに、構成が複雑になる。本研究では、勾配降下計算のための、より効率的な証明アーキテクチャを提案する。提案方式では、非線形計算にLUTに基づく手法を用い、残りの線形計算をストリーミング形式で処理する。本発表では、プロトコルの設計、順伝播および逆伝播との統合、ならびに量子化ニューラルネットワークへの実装について述べる。 | ||||||
| MUTABAZI PATRICK | D, 中間発表 | サイバーレジリエンス構成学 | 門林 雄基, | 笠原 正治, | 林 優一, | 妙中 雄三 |
|
title: RPS:Responsibility-Preserving Security for Zero-Trust Digital Systems
abstract: Modern digital security validates credentials, enforces access policies, and assesses device/session posture, yet failures still occur after access succeeds. Credentials may survive SIM-swap or recovery reassignment, sessions may persist after control changes, authorized requests may execute unintended semantics, and audit trails may identify accounts without preserving agency for fair accountability.This research introduces Responsibility-Preserving Security (RPS), a framework that treats identity, actor continuity, intent, authorization, execution, consequence, and accountability as one security chain. Its central claim is that credential validity, possession, and policy permission are necessary but insufficient for consequential digital action; systems must preserve evidence that actions remain attributable, express verified intent, execute authorized semantics, and support accountability under threat assumptions.RPS is developed through four contributions: telecom-anchored Zero-Trust identity and access management, Actor Integrity, Proof-of-Intent Authorization, and ZT-TOTP. Together, they show that modern security should protect not merely access, but legitimate responsibility across consequential digital actions. language of the presentation: English | ||||||
| NDALAMA FESTUS EDWARD （安本先生移動後開始） | D, 中間発表 | サイバーレジリエンス構成学 | 門林 雄基, | 安本 慶一, | 林 優一, | 妙中 雄三 |
|
title: Towards Secure Continuous Authentication for Internet Banking Using Adaptive Multimodal Behavioral Biometrics
abstract: Internet banking has become an essential financial service, but conventional authentication mechanisms such as passwords, PINs, and one-time passwords authenticate users only at login and cannot protect active sessions from threats such as session hijacking, impersonation, and account takeover. Continuous authentication based on behavioral biometrics offers a promising alternative by continuously verifying user identity through behavioral characteristics, including keystroke dynamics, mouse movements, touch interactions, and tap patterns. However, behavioral traits evolve over time due to device changes, environmental conditions, user habits, and adversarial manipulation, leading to behavioral concept drift that degrades authentication performance. This research investigates adaptive continuous authentication for Internet banking using multimodal behavioral biometrics. The work first introduces KeystrokeMouse-TZ, a context-rich behavioral biometrics dataset collected from realistic Internet banking interactions. Building on this foundation, the DABBiT framework is proposed to detect and adapt to natural and adversarial behavioral drift during continuous authentication. Finally, a security-aware adaptation governance mechanism, the Drift-Response Decision Engine (DRDE), is developed to control model adaptation using drift evidence, threat indicators, contextual information, and modality reliability, thereby preventing unsafe learning under adversarial conditions. Together, these contributions establish a comprehensive framework for secure, adaptive, and drift-resilient continuous authentication that improves the reliability, usability, and security of Internet banking systems across heterogeneous desktop and mobile environments. language of the presentation: English | ||||||
| 倉本 真菜 | M, 2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, | 西山 智弘 | |
|
title: Rewriting Medical Research Texts for Non-Experts Using Large Language Models
abstract: Medical research is highly specialized, making it difficult for patients and the general public to understand the latest research findings. Conventional approaches have focused on plain language rewriting, in which technical terms are replaced with easier expressions. However, this approach alone cannot compensate for the lack of background knowledge or provide omitted contextual information. In this study, we define rewriting for non-experts as an approach that not only simplifies technical expressions but also supplements the background and significance of medical research. We propose a rewriting method using large language models (Large Language Model: LLM). To reproduce the rewriting performed by professional medical writers, we constructed a dataset consisting of simulated medical research abstracts and corresponding rewritten texts created by the writers. We then compared multiple rewriting methods, including methods based on the writing guidelines used by the medical writers during dataset construction and Few-shot examples, using three LLMs. The results showed that a single-stage rewriting method combining the writing guidelines with Few-shot examples achieved the highest semantic similarity to the rewritten texts created by the medical writers. These findings suggest that, for rewriting for non-experts, directly generating rewritten texts based on writing guidelines and concrete examples is more effective than performing plain language rewriting in multiple stages. language of the presentation: Japanese 発表題目: 大規模言語モデルを用いた医学研究テキストのつたわる化 発表概要: 医学研究は高度に専門的であるため、患者や一般読者が最新の研究成果を理解することは容易ではない。そのため、従来は専門用語を易しい言葉に置き換える「平易化」が行われてきたが、それだけでは背景知識の不足や重要な情報の欠落を補えないという課題があった。そこで本研究では、平易化に加えて研究の背景や意義を補足する書き換えを「つたわる化」と定義し、大規模言語モデル（Large Language Model: LLM）を用いた手法を提案する。メディカルライターによるつたわる化の再現を目的として、模擬医学研究概要とメディカルライターが作成したつたわる化テキストからなるデータセットを構築した。さらに、データセット作成時にメディカルライターが用いたガイドラインやFew-shotを組み込んだ複数の書き換え手法を、3種類のLLMを用いて比較した。その結果、メディカルライターがデータ作成時に用いたガイドラインとFew-shotを併用した単一段階の書き換え手法が、メディカルライターが作成したつたわる化テキストとの意味的類似性において最も高い評価を示した。以上より、つたわる化では平易化を段階的に行うよりも、ガイドラインと具体例に基づいて直接生成する手法が有効であることが示された。 | |||||||
| 氷見 和 | M, 2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, | 西山 智弘 | |
|
title: Clinical Event Extraction from Japanese Electronic Health Record Documents Using Synthetic Data
abstract: Electronic health record documents contain valuable clinical information, including diagnoses, examinations, and treatments, in unstructured free-text form. However, obtaining sufficient training data for information extraction models is challenging because of restrictions on sharing clinical data and the high cost of expert annotation. In this study, we used a large language model (LLM) to transform publicly available lung cancer case reports into Japanese discharge summaries and progress notes and automatically annotate 37 clinical event items, resulting in 4,366 synthetic training examples. We then fine-tuned a small language model (SLM) using these data and evaluated it on 96 physician-authored and manually annotated simulated progress notes. The micro-F1 score was 0.261 in the zero-shot setting, 0.682 when using synthetic data alone, 0.767 when using manually annotated data alone, and 0.800 when the model was first trained on synthetic data and subsequently fine-tuned on manually annotated data. These results demonstrate that synthetic data derived from publicly available case reports can complement limited manually annotated data and support the development of clinical event extraction models for privacy-sensitive and resource-constrained health care settings. language of the presentation: Japanese 発表題目: 合成データを用いた日本語電子カルテ文書の構造化 発表概要: 電子カルテ文書には診断，検査，治療などの臨床情報が自由記述で蓄積されているが，医療データの共有制約や専門家によるアノテーションコストのため，情報抽出モデルの学習データ確保は困難である．本研究では，公開された肺がん症例報告を大規模言語モデル（LLM）により日本語の退院サマリと診療録へ変換し，37項目の臨床イベントを自動付与することで，4,366件の合成学習データを構築した．これを用いてSmall Language Model（SLM）をファインチューニングし，医師が作成・注釈した96件の模擬診療録で評価した．その結果，micro-F1はZero-shotの0.261に対し，合成データのみで0.682，人手注釈データのみで0.767，合成データによる学習後に人手注釈データで追加学習した場合に0.800となった．以上より，公開症例報告由来の合成データは，限られた人手注釈データを補完し，プライバシーと計算資源に制約のある医療環境に向けた臨床イベント抽出モデルの構築に有効であることが示された． | |||||||
| 田代 勇希 | M, 2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, | 西山 智弘 | |
|
title: MedNormJ: A Dataset for Medical Concept Normalization in Japanese Clinical Documents
abstract: Medical concept normalization in clinical text is a fundamental technology for the secondary use of clinical data. However, constructing annotated resources for this task is challenging because annotation is both expertise-intensive and methodologically complex. As a result, a standard evaluation dataset for Japanese has yet to be established. In this study, we introduce a Japanese dataset for medical concept normalization, MedNormJ , which will be publicly available. The dataset consists of 397 pairs of medical expressions and their corresponding normalized disease names, manually curated from 96 medical documents, including case reports and radiology reports. Furthermore, we conduct comparative experiments using existing normalization approaches to benchmark their performance on this dataset in terms of both accuracy and computational efficiency. Through these experiments, we clarify the present performance level and identify remaining challenges specific to Japanese medical concept normalization. language of the presentation: Japanese 発表題目: MedNormJ: 日本語臨床文書における医療用語正規化のためのデータセット 発表概要: 臨床テキストにおける医療用語正規化（medical concept normalization）は、臨床データの二次利用を支える基盤技術である。 しかし、このタスクのための注釈付き資源の構築は、専門知識を強く要するだけでなく、方法論的にも複雑であるため困難である。その結果、日本語における標準的な評価データセットはいまだ確立されていない。 本研究では、公開予定の日本語医学概念正規化データセットである MedNormJ を提案する。 本データセットは、症例報告および放射線診断報告書を含む96件の医学文書から人手で整理した、医学表現と対応する正規化疾患名の397対から構成される。 さらに、既存の正規化手法を用いた比較実験を行い、本データセットにおける精度および計算効率の両面から性能を評価する。 これらの実験を通じて、現在の性能を明らかにし、日本語医学概念正規化に固有の残された課題を特定する。 | |||||||
| 辻 航平 | D, 中間発表 | 自然言語処理学 | 渡辺 太郎, | 荒牧 英治, | 上垣外 英剛, | 坂井 優介, | 平岡 達也 |
|
title: Flexible LLM Vocabulary Design with Context-Dependent Vocabulary Sharing
abstract: Realistic LLMs must represent all languages within a finite vocabulary. Therefore, English-dominant LLMs have smaller vocabularies for non-Latin script languages like Japanese. This leads to inefficient processing and limits performance gains. However, simple vocabulary expansion or replacement methods face challenges such as increased computational costs and performance degradation. To address these issues, we introduce shared embedding spaces where multiple subwords share the same embedding. This approach expands the vocabulary size without increasing the embedding space. It minimizes computational inefficiency and performance degradation while resolving the traditional challenges. language of the presentation: Japanese 発表題目:文脈に依存した語彙共有による柔軟なLLM語彙設計 発表概要:現実的なLLMでは，有限の語彙であらゆる言語を表現する必要がある．そのため，英語が主のLLMでは日本語などの非ラテン文字体系言語の語彙サイズが小さくなり，処理が非効率かつ性能が上がりづらくなる．しかし，単純な語彙拡張や語彙置換には，計算負荷や性能低下などの課題がある．そこで本研究では，複数のサブワードが同じ埋め込みを使う埋め込み空間の共有を導入し，語彙サイズを広げることなく，埋め込み空間を小さくすることで，計算効率や性能低下を最低限に抑えながら従来の課題を解決する． | |||||||
| FAKIH ABDEL HAMID | M, 2回目発表 | サイバーレジリエンス構成学 | 門林 雄基, | 安本 慶一, | 妙中 雄三 | ||
|
title: Predwen: A Culturally Responsive, AI-Assisted Framework for Cyber Hygiene and Privacy Awareness among Japanese Non-IT Students
abstract: Predwen is a comparative cyber hygiene and privacy awareness project designed for Japanese non-IT students. The project evaluates whether a culturally contextualized, AI-assisted, and meaningfully gamified intervention can improve learning, engagement, intention, and delayed protective behavior compared with a generic gamified version without AI. Both experimental arms follow the ADDIE instructional design model and use the RECIPE meaningful gamification framework. The generic arm provides an individual, culture-neutral, CTF-light learning-by-doing pathway, while the localized arm extends the same structure through ADDIE+3I by adding Japanese contextualization, collectivist introspection prompts, and controlled AI-supported guidance. The platform is implemented as a privacy-preserving controlled online intervention using anonymous participant codes, pre/post testing, scenario-based tasks, feedback, an end questionnaire, and a J+7 behavioral follow-up. The study does not compare a passive awareness text against an interactive tool; both arms remain interactive and gamified to keep the comparison methodologically fair. Instead, the A/B design tests whether the localized AI-assisted package provides additional value over the generic interactive baseline. Outcomes are assessed through six metrics covering knowledge/confidence gain, engagement, resonance, intention, real adoption, and perceived AI help. Predwen therefore contributes a transferable framework for designing, localizing, supporting, and evaluating cyber hygiene education while preserving anonymity, minimizing data exposure, and avoiding competitive or punitive gamification. language of the presentation: English | |||||||
| 髙木 皐成 | M, 2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, | Sakriani Sakti, | 諏訪 博彦, | 佐々木 航 | |
|
title: Adaptive LLM-Generated Feedback for Calibrating Users’ Awareness of Smartphone Use
abstract: Smartphone use consists of brief and fragmented interactions, which can create discrepancies between users’ self-reported usage time and device logs. This study treats these discrepancies as a calibration problem distinct from reducing smartphone use and aims to support users in accurately understanding their own behavior. In a preliminary study with eight participants, natural-language feedback was consistently perceived as easier to understand, whereas numerical feedback provided a stronger sense of making one’s own judgment. Participants also differed in how they perceived evaluative or directive expressions. Although these findings do not demonstrate the effectiveness of personalization, they motivate an investigation of whether suitable feedback styles can be learned from individual responses. We propose a method that uses estimation errors as learning signals to sequentially select the style of LLM-generated messages presented on a smartwatch. The method integrates GAMBITTS, which learns from the features of the messages actually generated, with HierTS, which shares population-level information while personalizing decisions from sparse observations. Message generation and selection are performed on device so that raw usage logs and generated messages are not transmitted externally. A field study will evaluate whether message presentation and adaptive selection reduce estimation errors and whether any improvement persists after feedback is withdrawn. language of the presentation: Japanese 発表題目: スマートフォン使用の自己認識較正を支援する適応的LLMフィードバック 発表概要: スマートフォンの使用は短時間かつ断片的に繰り返されるため、利用者による使用時間の自己申告と端末ログにはズレが生じる。 本研究は、このズレを使用量の削減とは異なる問題として捉え、利用者が自分の使用状況を把握して判断するための自己認識の較正を支援する。 8名の予備実験では、自然言語を含む表示は理解しやすい一方、数値中心の表示は自分で判断する感覚が強く、評価感や押し付け感の受け止め方には個人差が見られた。 この結果は個人適応の効果を示すものではないが、利用者の反応から適切な伝え方を学ぶ必要性を検証する動機となった。 そこで、自己見積もり誤差を学習信号とし、スマートウォッチに提示するLLM生成メッセージの表現スタイルを利用者ごとに逐次選択する手法を提案する。 提案手法は、実際の生成文の特徴から学ぶGAMBITTSと、参加者間で全体傾向を共有しながら少数データで個人化するHierTSを統合する。 生成と選択は端末上で実行し、生の利用ログと生成文を端末外へ送らない。 今後、メッセージ提示と個人適応が見積もり誤差を縮小するか、提示停止後にも改善が残るかをフィールド実験で評価する。 | |||||||
| TANUSA YANISA | M, 2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, | Sakriani Sakti, | 諏訪 博彦, | 松井 智一 | |
|
title: In-the-wild Cross-Home Acoustic Activity Recognition via Home-alignment Contrastive Learning with CLAP Features
abstract: Human Activity Recognition supports healthcare, assisted living, and home monitoring. While cameras depend on light conditions and wearable sensors may be inconvenient, acoustic sensing can recognize activities without direct visual observation. However, real-home environments introduce variations in background noise, room conditions, and user behavior. This study proposes an in-the-wild cross-home acoustic activity recognition framework using Contrastive Language-Audio Pretraining (CLAP) embeddings and an MLP-BiGRU network with home-alignment supervised contrastive learning. Five-second audio segments are encoded by CLAP on a home server for privacy preserved, while the contrastive loss brings same-activity samples closer across different homes. Using leave-one-home-out evaluation across six homes, the proposed method achieved a mean F1-score of 60.27%, compared with 24.98% for the MLP baseline. Performance across homes ranged from 52.60% to 73.08%. However, activities with limited samples or variable sounds, such as washing, remain challenging. Future work will incorporate additional sensing modalities and natural-language activity descriptions to improve semantic understanding and generalization. language of the presentation: English | |||||||
| 長尾 洸吉 | M, 2回目発表 | ヒューマンＡＩインタラクション | Sakriani Sakti, | 清川 清, | 大内 啓樹, | Faisal Mehmood, | Bagus Tris Atmaja |
|
title: When Prejudice and Empathy Co-exist: Observers’ Attitudes Towards Conversational Robots abstract: Prejudice against conversational robots predicts negative attitudes towards them and it operates through empathy. However, whether the nature of operation of such a mechanism remains same across different roles i.e., interlocutor or observer? Such a research question is required to be addressed. In this preliminary research, we addressed such a research question through a subjective–evaluation based experiment. It was revealed that prejudice against conversational robots and negative attitudes towards them are linked with each other through empathic mechanism. However, the nature of operation of mechanism appears to differ from across roles. We explained the observed effects, highlighted the limitations of our findings, and concluded with future implications for human–robot interaction domain. language of the presentation: Japanese | |||||||