Seoul National University has announced that a research paper by Professor Eom Dae-ho and his team from the Department of Electronic Engineering has been accepted for presentation at the CVPR 2026, one of the most prestigious international conferences in the field of artificial intelligence and computer vision. This conference, organized by the IEEE/CVF, will be held in Denver, USA, from June 3 to June 7, 2026. The paper titled 'When CLIP Sees More, It Fights Back Harder: Multi-View Guided Adaptive Counterattacks for Test-Time Adversarial Robustness' presents a novel defense technique aimed at improving the vulnerability of large vision-language models (CLIP) to adversarial attacks that cleverly manipulate images.

In their research, the team proposed a method that applies different transformations to input images to generate multiple views, which then allows them to adaptively adjust the strength of counterattacks based on the estimated damage level for each view. This approach effectively addresses the issue where existing methods experience a drastic drop in performance under strong attack conditions. They achieved state-of-the-art adversarial robustness across 20 experimental datasets, demonstrating fast inference speed and low memory usage without the need for additional training or parameter tuning.

Professor Eom noted that this research showcases the potential for vision-language models to effectively mitigate security threats in real-world scenarios at test time, emphasizing that robust stability can be achieved without additional training data or model adjustments. This work is expected to have significant implications for various AI applications where safety is critical, such as autonomous driving, medical imaging analysis, and industrial vision systems.