- Research team led by Professor Daeho Um of the School of Electrical and Computer Engineering develops a multimodal AI technique robust against image manipulation attacks, following another major achievement at CVPR 2026

- Enhances the reliability of large vision-language models through stability and suitability metrics

  The University of Seoul announced that a paper by the research team led by Professor Daeho Um of the School of Electrical and Computer Engineering has been accepted to the International Conference on Machine Learning (ICML 2026), one of the world’s most prestigious international conferences in the field of artificial intelligence (AI).

  ICML, together with Neural Information Processing Systems (NeurIPS) and International Conference on Learning Representations (ICLR), is regarded as one of the world’s top three machine learning conferences and a leading venue driving advances in artificial intelligence (AI). Marking its 43rd edition, ICML 2026 is scheduled to take place from July 6 to July 11 at the COEX Convention & Exhibition Center in Seoul. This will be the first time ICML has been held in person in Korea.

▶ Proposed Method: SS-TPT

  The accepted paper, titled “SS-TPT: Stability and Suitability-Guided Test-Time Prompt Tuning for Adversarially Robust Vision-Language Models,” proposes a novel defense method that effectively addresses the vulnerability of large vision-language models (CLIP) to image manipulation attacks. The method generates multiple views of input images through diverse transformations and evaluates two newly defined metrics—stability and suitability—to assess the reliability of each view. Based on these metrics, the framework places greater emphasis on more reliable views during inference. The proposed approach achieves state-of-the-art adversarial robustness and significantly improves the reliability of large vision-language models without modifying their internal parameters.

▶ Professor Sunoh Kim of Dankook University (left, first author), Professor Daeho Um of the University of Seoul (right, corresponding author)

  This study was conducted as a collaborative research project between Professor Sunoh Kim of Dankook University (first author) and Professor Daeho Um of the University of Seoul (corresponding author).

  Professor Daeho Um stated, “Following our work presented at CVPR 2026, this study proposes a new direction for enhancing the robustness of vision-language models against adversarial attacks.” He added, “Through continued collaboration with Professor Sunoh Kim, we plan to continue advancing research on the reliability of vision-language models and contribute to enhancing the trustworthiness of AI in everyday life.”