The Ministry of Science and ICT is set to develop a foundation model specialized in cybersecurity, following its work on general-purpose artificial intelligence (AI) models. This initiative aims to enhance security expertise based on domestic AI models and diversify the model ecosystem.
However, with significant investments already being made in general-purpose models, proving the performance and efficiency of specialized models is seen as crucial, given the additional computing resources required.
According to the Ministry, the application period for the 'Specialized AI Foundation Model Development' project closed on August 26. SK Telecom and Naver Cloud have formed separate consortiums to support the initiative.
The Ministry plans to select a final contractor after evaluations and will provide 256 NVIDIA GPU B200 units for 10 months. The goal is to develop a foundation model tailored for cybersecurity based on domestic AI technologies and models. It is also possible to enhance security expertise using existing domestic foundation models.
SK Telecom aims to develop a security-focused model based on its proprietary AI foundation model. Naver Cloud has not disclosed specific details about its foundational model or development approach.
The Ministry's decision to pursue the development of specialized models stems from the belief that general large language models (LLMs) have limitations in performing specialized security tasks such as malware analysis and threat detection. The initiative seeks to improve the AI capabilities in the domestic security industry by training on security data and specialized knowledge.
However, there are concerns regarding the costs and complexity of developing specialized models, which can vary significantly based on the development approach.
Cho Sung-bae, a professor of computer engineering at Yonsei University, noted, "Creating a foundation model from scratch has the advantage of being designed in the desired direction, but it requires vast amounts of data, time, and costs." He added, "Using 256 B200 units over 10 months to create a foundation model may be a tight timeline." He suggested that a more feasible approach might be to adjust existing open weights or open-source models toward a security specialization rather than starting from scratch.
Securing performance appropriate for specialized models is also a challenge. Choi Jae-sik, a distinguished professor at KAIST, explained, "Specialized models have relatively less data compared to general models and need to incorporate specialized knowledge. They also require a higher level of accuracy in practical applications, producing results without hallucinations."
Despite the challenges in developing specialized models, there is an assessment that nurturing both general models and specialized models in various fields could enhance the competitiveness of the domestic AI ecosystem in the long run.
Jung Jun-hwa, a legislative researcher at the National Assembly's Legislative Research Service, stated, "It is worth reconsidering whether concentrating support on a small elite team is appropriate for enhancing the overall AI resilience of the nation in times of crisis. Both top-down approaches that derive services from foundation models and bottom-up approaches that nurture small specialized AI models in various fields should be considered."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.