Dongsu Song

About Me

Hello! I am Dongsu Song. Thank you for your interest. I received my M.S. degree in Artificial Intelligence from Korea Aerospace University in 2025, and I was fortunate to conduct my master’s research under the supervision of Prof. Jay Hoon Jung.

My research focuses on the reliability of vision models: finding where they fail, understanding why, and making them more robust. I began with black-box sparse attacks that expose a model’s weaknesses by perturbing only a few pixels under limited query access (NeurIPS 2024, ECCV 2026). I have since moved from pixel-level noise to natural semantic changes, such as a shift in lighting to dusk, that degrade models across architectures. Going forward, I want to take this failure-discovery work from single images to video and 3D scene understanding, and study how visual failures propagate from perception to action.

Research Interests

• Adversarial Robustness: Finding and understanding adversarial vulnerabilities of vision systems, building on black-box sparse attacks and extending to video, 3D, multimodal foundation models, and vision-language-action (VLA) models
• Generative Failure Discovery: Actively searching for semantic edge cases with generative models, understanding why vision models fail on them, and using these insights to make models more robust
• Visual Privacy: Repurposing adversarial perturbations to protect images and videos against unauthorized generative editing, image animation, and 3D/4D reconstruction

News

• [2026.06.18] Paper accepted at ECCV 2026!
• [2026.06.02] Selected as a Reviewer for NLDL 2027.
• [2025.08] Invited presentation at KCCV 2025!
• [2025.06] Selected as a Reviewer for NeurIPS 2025.
• [2024.11] Paper accepted at NeurIPS 2024!
• [2023.10] Paper accepted at Journal of KIISE!

Selected Papers