Naver Cloud and LG will jointly develop AI models specialized for security attacks and defense. Their goal is to develop world-class cybersecurity AI by training the two models on the same security data through different methods and cross-validating them to boost both attack and defense capabilities.
According to the Naver Cloud-LG CNS consortium on the 10th, both companies will begin developing security-specialized AI in earnest on the 18th. The consortium will carry out the project divided into five teams: Data Acquisition Team, Data Preprocessing Team, Foundation Model (FM) Development Team, Evaluation and Verification Team, and Policy and Commercialization Team. A total of 33 institutions, including Naver Cloud, are participating.
Naver Cloud will first strengthen defense capabilities based on 'HyperCLOVA X.' LG will enhance attack capabilities using 'EXAONE' from LG AI Research. HyperCLOVA X will inject security data starting from the pre-training stage, while EXAONE will apply continued pre-training (CPT) to its existing foundation model.
The ultimate goal for both companies is a 700 billion (700B) parameter Mixture of Experts (MoE) model. They plan to first create models specialized in attack and defense, respectively, and then improve performance on both sides through cross-validation by repeating mutual attacks and defenses.
For performance verification benchmarks, 'CyberGym' will be used for the attack sector and 'ExCyTIn-Bench' for the defense sector. The consortium set a goal to achieve 100% of the performance level of leading global models.
Jeon Eun-sook, head of the Security Business Division at LG CNS, said, “Although it may be difficult to raise the performance of all general-purpose features at once, I believe we can reach global standards if we focus on security,” expressing her challenge to achieve the world's No. 1 spot based on benchmark standards.
Model training will utilize 831.59TB of security data across 126 items currently secured from 21 institutions. The consortium plans to refine and de-identify the data before putting it into model training.
The consortium plans to showcase initial results utilizing security-specialized AI within the year. Afterward, it will refine the models through an interim evaluation in February next year and complete the project in July next year. The two developed models are planned to be released as open source for commercial use. They will also be provided in the form of APIs, SDKs, and security agents so that domestic security companies can apply them to existing products and services.
In the future, the security-specialized AI models are expected to be provided as Software-as-a-Service (SaaS)-based AI Application Programming Interfaces (APIs), Software Development Kits (SDKs), and security agents. The idea is to enable domestic security companies to utilize them for existing products and services.
The developed models will be demonstrated across seven key sectors: electric power, finance, science and technology, telecommunications, semiconductors, defense industry, and aerospace. Performance and usability will also be verified in actual industrial environments, such as closed networks and on-premises setups.
Integration with domestic AI semiconductors will also be pursued. The plan is to run the trained security models on domestic AI semiconductors to verify inference performance, and later expand into a sovereign AI security framework that encompasses models, cloud, and domestic semiconductors.
They will also expand into overseas markets by leveraging the existing business networks of Naver Cloud and LG CNS. They plan to link security-specialized AI to Naver Cloud's overseas 'AI Factory' business and spread it through LG CNS's local customer and partner networks across 13 overseas bases.
Kim Jin-hwee, senior vice president of Naver Cloud, said, “As cyberattacks using AI continue 365 days a year and attack traffic surges rapidly, we have entered an era where defenders' use of AI has also become essential,” adding, “We plan to cooperate with participating companies to develop models that can be practically utilized in the public and private sectors.”
