Researchers at UNIST has earned international recognition for developing a method to detect and mitigate hidden malicious triggers in artificial intelligence systems—an emerging threat to the reliability of large language models (LLMs).

Led by Professor Saerom Park of the Department of Industrial Engineering and the Graduate School of Artificial Intelligence, and Professor Sung Whan Yoon of the Graduate School of Artificial Intelligence and the Department of Electrical Engineering, the team placed second in the Anti-Backdoor (Anti-BAD) Challenge at the IEEE Conference on Secure and Trustworthy Machine Learning (SaTML 2026), held in Munich, Germany, from March 23 to 25, 2026.

Backdoor attacks embed hidden signals into AI models during training, causing them to produce unintended outputs when specific inputs—known as triggers—are encountered. Because these models otherwise perform as expected, these vulnerabilities present a persistent challenge for the safe deployment of AI systems. 

The competition challenged participants to develop methods, capable of reducing the impact of such triggers across a range of applications, including text generation, classification, and multilingual tasks. In response, the UNIST team proposed a unified framework designed to operate effectively across these varied settings. 

Their approach integrates model quantization, model merging, outlier parameter detection, and confidence calibration. Together, these techniques enable the identification and suppression of anomalous behaviors while preserving overall model performance. The methods does not rely on prior knowledge of attack patterns, making it applicable to a broad class of models and use cases. 

Contributors to the study include JiEun Yun and KiWan Kwon of the Department of Industrial Engineering, and SeungBum Ha of the Graduate School of Artificial Intelligence. 

"Even in the absence of prior information about the attack methods, it is possible to meaningfully reduce backdoor risks in large language models,” said JiEun Yun. "This work is a step toward ensuring that AI systems can be deployed with greater confidence." 

The IEEE SaTML is a leading international forum dedicated to advancing research on the security, robustness, and trustworthiness of machine learning systems. The annual competition serves as a benchmark for emerging approaches to safeguarding next-generation AI technologies.