Futa Waseda
Project Assistant Professor | Trustworthy AI: Robustness, Reliability, and VLM Defense
National Institute of Informatics (NII)
About Me
I study Trustworthy AI, focusing on adversarial robustness and reliable generalization from mechanism to deployment:
- I analyze why models fail under adversarial perturbation (WACV'23).
- I design robust training and defense methods for computer vision (ICLR'25, ICIP'24).
- I extend these ideas to vision-language systems (ECCV'26, WACV'26, ACMMM'25).
- I also work on complementary reliability and security topics, including model IP protection (ACL'25 Main) and post-hoc calibration (ICML'23).
I have collaborated widely across academia and industry through joint research and internships, including work with TUM, NII, NEC, CyberAgent AI Lab, SB Intuitions, and Turing.
Interests
- Trustworthy AI
- Adversarial Robustness
- Robust Generalization
- Vision-Language Model Reliability
Education
PhD in Informatics, 2026
The University of Tokyo
MS in Informatics, 2023
The University of Tokyo
Exchange Student, 2022
Technical University of Munich
BEng in Systems Innovation, 2020
The University of Tokyo
News
- 2026.06: One 1st-authored paper accepted at ECCV 2026.
- 2026.02: One co-authored paper accepted at CVPR 2026.
- 2026.01: One co-authored paper accepted at ICLR 2026.
- 2026.03: Received PhD in Informatics from The University of Tokyo.
- 2025.11: One 1st-authored paper accepted at WACV 2026.
- 2025.09: Received NII Inose Outstanding Student Award, 2025.
- 2025.07: One 1st-authored paper accepted at ACMMM 2025.
- 2025.05: One 1st-authored paper accepted at ACL 2025 Main.
- 2025.01: One 1st-authored paper accepted at ICLR 2025.
- 2024.09: One 1st-authored paper accepted at ICIP 2024.
- 2024.05: One 1st-authored paper accepted at MIRU 2024 Oral.
- 2023.09: Selected for research fellowship JST DC2.
- 2023.07: Received a research grant of 1 million yen from AIP Challenge Program, JST.
- 2023.03: Selected for research fellowship JST SPRING GX.
- 2023.04: One 1st-authored paper accepted at ICML 2023.
- 2022.10: One 1st-authored paper accepted at WACV 2023.
Recent Publications
Research strengths
- Defenses for visual and vision-language models under adversarial settings.
- Reliable generalization via calibration and robustness analysis.
- Practical security themes, including model IP protection.
See all publications.
ECCV 2026
September 2026
Read or Ignore? A Unified Benchmark for Typographic-Attack Robustness and Text Recognition in Vision-Language Models
- Introduces RIO-Bench, the first benchmark unifying typographic-attack robustness and text recognition evaluation for VLMs.
- Shows that strong VLMs and existing defenses fail to balance robustness against typographic attacks and text-reading capability.
- Proposes a data-driven defense enabling adaptive selective text use, outperforming prior non-adaptive approaches.
CVPR 2026
June 2026
Text-Printed Image: Bridging the Image-Text Modality Gap for Text-centric Training of Large Vision-Language Models
- Proposes Text-Printed Image (TPI), a method that synthesizes training images by rendering text directly on a plain canvas, enabling text-centric VLM training without costly image generation.
- Bridges the image-text modality gap for large vision-language models using only textual data, dramatically reducing training cost.
- Demonstrates that TPI outperforms diffusion-based synthetic images across multiple benchmarks while better preserving semantic accuracy.
ICLR 2026
May 2026
Understanding Sensitivity of Differential Attention through the Lens of Adversarial Robustness
- Reveals that Differential Attention (DA), despite improving discriminative focus by suppressing redundant context, paradoxically increases adversarial sensitivity via negative gradient alignment.
- Provides theoretical analysis and empirical validation across Vision Transformers and CLIP models, showing consistently higher attack success rates in DA-equipped models.
- Identifies a fundamental trade-off between clean-input discriminability and adversarial robustness in DA, with depth-dependent protection patterns against small perturbations.
WACV 2026
March 2026
Multimodal Adversarial Defense for Vision-Language Models by Leveraging One-To-Many Relationships
- Targets robust defense for VLMs under multimodal adversarial attacks.
- Leverages one-to-many image-text relationships during adversarial training.
- Improves robustness while preserving clean-task performance.
APSIPA ASC 2025
December 2025
Uncolorable Examples: Preventing Unauthorized AI Colorization via Perception-Aware Chroma-Restrictive Perturbation
- Proposes PAChroma, which optimizes imperceptible perturbations with a Laplacian filter to generate Uncolorable Examples that resist unauthorized AI colorization.
- Ensures visual indistinguishability from clean images while achieving strong transferability and robustness across diverse AI colorization models.
- Protects monochrome content (manga, historical films) from unauthorized AI colorization and commercial resale, addressing a practical copyright protection need.
ACMMM 2025
October 2025
Quality Text, Robust Vision: The Role of Language in Enhancing Visual Robustness of Vision-Language Models
- Proposed QT-AFT to improve VLM robustness with high-quality text supervision.
- Reduced class-overfitting while improving robustness to text-guided attacks.
- Achieved strong zero-shot robustness and accuracy across 16 datasets.
ACL 2025 Main
July 2025
MergePrint: Merge-Resistant Fingerprints for Robust Black-box Ownership Verification of Large Language Models
- Proposed MergePrint for merge-resistant fingerprinting of LLM ownership.
- Enabled black-box verification without requiring model internals.
- Preserved fingerprint detectability after model merging with limited utility loss.
ICLR 2025
April 2025
Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off
- Revisited invariance regularization to improve the robustness-accuracy trade-off.
- Proposed AR-AT with asymmetric loss, stop-gradient, predictor, and split-BN.
- Learned adversarially invariant yet discriminative representations more effectively.
ICIP 2024
October 2024
Defending Against Physical Adversarial Patch Attacks on Infrared Human Detection
- First defense study for physical adversarial patches in infrared human detection.
- Proposed POD, an efficient patch-aware training and detection strategy.
- Achieved strong robustness to unseen patch attacks while improving clean precision.
ICML 2023
July 2023
Beyond In-Domain Scenarios: Robust Density-Aware Calibration
- Proposed DAC, a KNN-based density-aware post-hoc calibration method.
- Improved uncertainty reliability under domain shift and OOD conditions.
- Maintained strong in-domain performance across diverse models and datasets.
Experience
Project Assistant Professor
National Institute of Informatics
Research focus: trustworthy AI, adversarial robustness, and vision-language model reliability.
Research Internship → Contractor
Turing
Research Internship (Apr 2025 – Mar 2026) → Research Contractor (Apr 2026 – present). Research focus: reliable vision-language models.
Research Internship
CyberAgent AI Lab
Research focus: deep learning, vision-language models, and adversarial robustness.
Research Internship
NEC Corporation
Research focus: deep learning, computer vision, adversarial robustness, and parameter-efficient training.
Technical Advisor
Madori LABO（まどりLABO）
Co-founding engineer at a startup building an AI-driven floor-plan design assistant.
Research Assistant
National Institute of Informatics
Research focus: deep learning, computer vision, and adversarial robustness.
Technical Advisor
Ollo inc.
Started as a machine learning engineer and now provide technical guidance from a researcher’s perspective.
Awards
NII Inose Outstanding Student Award (NII猪瀬優秀学生賞), 2025
Received the NII Inose Outstanding Student Award (NII猪瀬優秀学生賞).
MIRU'24 Student Encouragement Award (MIRU'24 学生奨励賞)
Received the MIRU 2024 Student Encouragement Award (MIRU'24 学生奨励賞).
Won the special prize at SAS analytics hackathon 2019.（SAS Institute Japan, The Analytics Hackathon 2019 特別賞）
In the contest, participants were given data and asked to construct machine learning system with high accuracy. (article url: https://enterprisezine.jp/article/detail/12209?p=2)
Won the first prize at MDS data science contest 2018.（MDSデータサイエンスコンテスト 優勝）
In the contest, participants were given big data and asked to perform value-generating analysis freely. Our group won the first prize and we were able to submit a paper. See the publication section.
Accomplishments
Summer School for Deep Generative Models 2020
Learned deep generative models from basics to state-of-the-art.
Chair for Global Consumer Intelligence (GCI 2018)
Learned how to utilize big data by machine learning technology.
Side Projects
Floor Plan App (間取り生成アプリ)
Floor Plan App is a web application that generates a floor plan from a given request.
Twitter Image Captioning
Made a model which outputs text from a image like human tweets, using Encoder-Decoder Model. Application of image captioning technique.
Oshaberi-Bot（おしゃべりぼっと）
My first twitter bot app. He learns japanese from his followers, by fitting retrieved data to Markov model.
trip map
Demo web application I made in school. In this app, you can clip the place you want to go in the future, find the shortest way to go through the chosen spots. I was responsible for front-end system using html, css, javascript.