| BORHAN ROWNAK | D, 中間発表 | サイバーレジリエンス構成学 | 門林 雄基, | 林 優一, | 藤本 大介, | 妙中 雄三 |
|
title: Next-Generation Security for IIoT: A Quantum Resistant Cryptographic Approach
abstract: Quantum computing is an emerging threat to the cryptographic mechanisms that currently protect information worldwide. Cryptographic methods that were once considered secure are now at risk of being compromised, creating an urgent need to migrate toward quantum-resistant cryptographic systems. However, this transition is complex, and many organiza- tions lack the in-house expertise required to navigate it without structured guidance. The challenge is even greater in IIoT in- frastructures, where the presence of resource-constrained devices, multiple architectural layers, legacy protocols, and diverse op- erational requirements makes the migration process significantly more difficult. Motivated by these challenges, this research proposed a PQC-based signcryption method and MIRAI Shield, a comprehensive framework designed to support IIoT-based industries in the PQC migration process. language of the presentation: English | ||||||
| 谷口 蒼馬 | M, 2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: CAPTCHA Authentication that Exploits the Weaknesses of AI — Triggering Safety Guardrails and Leveraging Vulnerabilities in Negation Processing of LLMs
abstract: CAPTCHA is a technology for distinguishing humans from bots, yet with the advancement of large language models (LLMs), conventional CAPTCHAs are now broken with high accuracy. Since it is LLMs that break CAPTCHAs, we believe that questions exploiting the weaknesses of LLMs can serve as an effective defense. This talk reports two approaches based on this idea. The first embeds phrases that trigger LLM safety guardrails into challenges to make bots refuse to answer. In preliminary experiments, however, most LLMs still answered questions about animal images even when such trigger sentences were embedded, and refusals were induced only in limited cases. On the other hand, such prompts can trigger terms-of-service enforcement on commercial AI services and lead to suspension of attackers' accounts, offering a secondary deterrent effect. The second exploits LLMs' vulnerability in processing negation: presenting true/false questions in negative or double-negative form lowers the accuracy of standard LLMs, while reasoning-oriented LLMs incur tens of times the computational cost of standard models. We will work toward realizing a new CAPTCHA that LLMs cannot answer, including by integrating the two approaches. language of the presentation: Japanese 発表題目: AIの弱点を逆手に取るCAPTCHA認証 —安全機構の作動誘発と否定表現処理の脆弱性の活用— 発表概要: CAPTCHA認証は人間とボットを判別する技術であるが、大規模言語モデル（LLM）の発展により従来のCAPTCHA認証は高精度で突破されるようになった。CAPTCHAを破るのがLLMである以上、LLMの弱点を突く問いこそが有効な防御になると考える。本発表では、この着想に基づく二つのアプローチを報告する。第一は、LLMの安全機構（ガードレール）を作動させる文言を設問に埋め込み、ボットに回答を拒否させる手法である。予備実験では、動物画像への質問に安全機構を作動させる文章を埋め込んでも多くのLLMが回答を返し、回答拒否の誘発は限定的であった。一方で、こうしたプロンプトは商用AIサービスにおける規約違反検知を誘発し、攻撃者のアカウント停止につながり得るという副次的な抑止効果が期待できる。第二は、LLMの否定表現処理の脆弱性を利用する手法であり、真偽判定問題を否定形・二重否定形で出題すると標準的なLLMの正答率が低下し、推論特化型でも標準型の数十倍の計算コストを要することを示した。今後は両者の統合などにより、LLMには回答できない新たなCAPTCHA認証の実現に向けて研究に取り組む。 | ||||||
| 柴田 日向 | M, 2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 | |
| 発表テーマ： 依存関係を用いたRustバイナリからのクレート復元手法 発表概要： Rustバイナリでは, 標準ライブラリや依存クレート由来コードが固有コードと混在し, 解析範囲の切り分けが難しい.本研究では, バイナリ中の残存痕跡からクレートを直接検出し, crates.ioの依存関係から未検出候補を推定する.11件のRust製OSSによる評価では, 直接検出は高いPrecisionを示した一方, 強いstrip条件ではRecallが低下した.依存推定は補助情報として利用できるが, 網羅的な復元には追加の手掛かりが必要である. Title: Crate Recovery from Rust Binaries Using Dependency Relationships Abstract: Rust binaries commonly contain standard-library and dependency-crate code mixed with application-specific code, making it difficult to narrow the analysis scope. This study directly detects crates from residual binary artifacts and infers additional candidates using dependency information from crates.io. An evaluation on 11 Rust open-source projects showed high precision for direct detection, while recall decreased under aggressive stripping. Dependency-based inference can provide supplementary information, but additional evidence is required for comprehensive crate recovery. language of the presentation: Japanese | ||||||