森が日本セキュリティ・マネジメント学会 第39回全国大会（統一テーマ「AI時代のリスクとセキュリティ・マネジメント」）において「フィジカルAIセキュリティ」の特別講演をしました。
日本セキュリティ・マネジメント学会第39回全国大会で特別講演
Author Archives: mori
A paper got accepted!
We are thrilled to announce that our conference paper entitled “Artistic Adversarial Examples: Graffiti-Based Adversarial Attacks on Traffic Sign Recognition” has been accepted for publication in The 28th International Conference on Information and Communications Security (ICICS 2026). This work is a fruit of our international collaboration with Politecnico di Milano. Congratulations to Zhuge-kun and kudos to the entire team!
In this work, we propose artistic adversarial examples — adversarial patches crafted to look like ordinary street graffiti so that they blend naturally into urban scenes. Combining a graffiti-trained generative model with a two-stage optimization that couples white-box latent optimization and black-box placement refinement, our attack degrades traffic sign detection while keeping the perturbations visually plausible. Extensive experiments across digital, simulated, and physical settings show that the generated graffiti substantially disrupts traffic sign recognition, achieving up to a 72.8% attack success rate with an average confidence drop of 65.8%.
Zhenghao Michele Zhuge, Go Tsuruoka, Zhihe Zhang, Stefano Longari, Lachlan Moore, Stefano Zanero, and Tatsuya Mori, "Artistic Adversarial Examples: Graffiti-Based Adversarial Attacks on Traffic Sign Recognition." In Proc. of The 28th International Conference on Information and Communications Security (ICICS 2026), Fukui, Japan, Oct 2026.
Guest talk by Prof. Zhiqiang Lin (Ohio State University)
On June 24, 2026, we were honored to host Prof. Zhiqiang Lin (Distinguished Professor of Engineering, and Director of the Institute for Cybersecurity and Digital Trust, The Ohio State University) at our laboratory on the Nishi-Waseda campus. Prof. Lin gave a research talk followed by an open discussion with our students and invited researchers.
Title: Validated Exploits, Not LLM Claims: Runtime-Grounded Security Agents for AI-Generated Code
Abstract: AI coding agents are rapidly changing how software is produced, but they also create a new assurance problem: generated code can be deployed faster than humans can review it. This talk examines how large language models can be used not merely to flag suspicious code, but to construct evidence-backed security findings for full-stack web applications. Prof. Lin presented BuckAI, an LLM-agent pipeline for web vulnerability discovery that combines code reasoning, static analysis, automated deployment, exploit generation, and runtime validation. The central design principle is that LLMs should propose hypotheses, while validators and runtime oracles determine whether those hypotheses correspond to real, exploitable behavior. The talk covered the challenges of building such a system — the lack of runnable benchmarks, the difficulty of deploying diverse web applications, and the unreliability of LLMs as final security judges — and concluded with lessons for building trustworthy security agents and open research directions in agentic AI for software security.
We thank Prof. Lin for the stimulating talk and the lively discussions with our group, and we look forward to future collaborations.
A paper got accepted!
We are thrilled to announce that our conference paper entitled “Detecting Phishing on Shared-Domain Hosting Services Using LLM-Based Contextual Mismatch Reasoning” has been accepted for publication in The 21st International Conference on Availability, Reliability and Security (ARES 2026). Congratulations to Inuzuka-kun and kudos to the entire team!
In this work, we propose an LLM-based phishing detection workflow tailored to shared-domain hosting services (SHSs) such as free website builders, where benign and malicious pages coexist under the same provider-managed domain and conventional URL- and domain-based features break down. Our three-stage pipeline — Brand Analyzer, Content Analyzer, and Judge — reasons about the contextual mismatch between the brand a page claims to represent and the sensitive actions it prompts. Evaluated across ten SHSs with both commercial (GPT-4.1-mini) and open-source (Qwen3-8B) models, our system achieves high recall, outperforms state-of-the-art baselines including FreePhish and PhishLLM, and provides natural-language rationales that explain each verdict.
Sho Inuzuka, Takaaki Toda, Daiki Chiba, and Tatsuya Mori, "Detecting Phishing on Shared-Domain Hosting Services Using LLM-Based Contextual Mismatch Reasoning." In Proc. of The 21st International Conference on Availability, Reliability and Security (ARES 2026), Linköping, Sweden, Aug 2026.
A paper got accepted!
We are thrilled to announce that our journal paper entitled “Comprehensive Evaluation and Defense against Adversarial Fog Attacks on LiDAR-Based Autonomous Driving” has been accepted for publication in the Journal of Information Processing (JIP). Congratulations to Tanaka-kun and kudos to the entire team!
This paper is an extended journal version of our ACM AsiaCCS 2025 paper. In this work, we present a comprehensive evaluation of Adversarial Fog Attacks (AFA), which exploit vulnerabilities of point cloud preprocessing filters in LiDAR-based autonomous driving. We systematically analyze how environmental and attack parameters affect the attack effectiveness, physically validate the attack across multiple LiDAR models, and investigate post-detection avoidance behaviors together with practical defense strategies. The paper will appear in Vol. 34 (September 2026).
Yuuna Tanaka, Kazuki Nomoto, Ryunosuke Kobayashi, Go Tsuruoka, and Tatsuya Mori, "Comprehensive Evaluation and Defense against Adversarial Fog Attacks on LiDAR-Based Autonomous Driving." Journal of Information Processing, Vol. 34, Sep 2026.