GIST Develops New Technology to Pinpoint Errors in Autonomous Driving AI

Focus on "latent space"... Apply LSAA attack

Gwangju Institute of Science and Technology (GIST) said on October 20 that a research team led by Professor Kim Seungjun of its Department of AI had developed a new adversarial attack technique, "LSAA (Latent Space-Aware Adversarial Attack)," that effectively identifies vulnerabilities in AI visual recognition.

AI Vulnerability Assessment Process Using LSAA. Provided by GIST.

An "adversarial attack" is a technique that introduces subtle changes to image data to induce AI errors. In real-world settings, it is used as a safety testing tool with characteristics of "penetration testing" to check in advance the conditions under which an AI system may make errors.

Self-driving cars use "semantic segmentation" AI technology, which identifies objects at the pixel level to distinguish roads, pedestrians, obstacles, and other elements. However, existing vulnerability testing methods design attacks based on similarities defined by humans, failing to fully reflect the AI's actual internal recognition mechanisms. They also have limitations in that they become more difficult to apply as the number of object categories increases.

The research team focused on the "latent space," an internal space where AI quantifies and organizes the features of objects. The technique identifies objects that AI classifies as similar and is therefore likely to confuse in the latent space, then introduces subtle changes. It also updates the targets of the attack in real time as the image changes, maximizing misrecognition.

In tests using urban driving footage, the AI's mean intersection over union (mIoU) for object regions fell sharply from 68.35% under normal conditions to 0.41% when the LSAA attack was applied. Experiments using a dataset containing 150 diverse object categories also confirmed the technique's general applicability, with the mIoU falling to 0.6% or less.

Professor Kim Seungjun said, "The answer to what confuses AI lies in the internal latent space formed during training. This technology will make a significant contribution to precisely diagnosing vulnerabilities in visual AI systems installed in self-driving cars and intelligent robots in advance and improving their safety."

Kim Gwangbin, a doctoral student, participated in the study as first author. Supported by the Ministry of Science and ICT and the Ministry of Land, Infrastructure and Transport, the research was presented recently at IROS 2026, a leading robotics conference held in Pittsburgh, United States.

This content was produced with the assistance of AI translation services.