Maintaining Customized Performance and Ensuring Safety for Personalized AI... KAIST Unveils Core Technology Development
Large Language Models (LLMs) are designed with "safety alignment" to prevent them from responding to users' dangerous requests. However, as more individuals and businesses are independently training AI models on their own documents and data to create personalized AIs, concerns are growing that the existing safety alignment could be compromised. Ensuring both customized performance and safety has thus emerged as a key challenge in the era of AI personalization. Now, a core AI technology that maintains customized performance while enhancing safety in the era of "personalized AI" has been developed domestically.
On July 15, KAIST announced that the research team led by Professor Changik Kim from the Department of Electrical and Electronic Engineering has developed the "Buffer-and-Reinforce" learning framework to address the issue of safety deterioration that occurs during the "fine-tuning" process, where LLMs such as ChatGPT are further trained using data from individuals and companies.
The research team focused on prior research indicating that the safety of customized AI does not significantly deteriorate, even when the AI is in a "jailbreak" state—meaning it responds to dangerous requests without rejection. Based on this, they devised a new approach.
The key innovation is not using the jailbreak state in actual services, but instead applying it temporarily during customized training using a buffering module called "BufferLoRA," and then removing it afterward.
Through this method, the team confirmed that an AI in the jailbreak state can effectively learn new work capabilities desired by the user without being influenced by dangerous information. The principle is that the AI continues to learn only the necessary knowledge, without additionally learning dangerous information.
Based on this principle, the research team developed a two-stage learning technique that uses "buffering" and "reinforcing" for safety.
First, the BufferLoRA module is temporarily applied to the AI, acting as a protective shield during customized training to prevent malicious data from directly affecting the main model. Once the customized training is completed, this module can be removed.
Next, the AI’s safety is enhanced again using the safety reinforcement module, "ReinforceLoRA." This is achieved through the mathematical method of "QR decomposition," which separates different information and selectively reflects only the necessary information. In this way, it retains the new functions learned by the user while allowing the AI to selectively enhance only its safety.
In summary, a temporary protective shield (BufferLoRA) is applied to the main AI to block direct impact from malicious data, allowing the AI to learn only the necessary work-related capabilities. After the shield is removed, the safety reinforcement module (ReinforceLoRA) further strengthens the AI’s safety alignment.
(From left) Wonjun Lee, PhD candidate; Changik Kim, Professor; Seokil Ham, PhD candidate; Jaehyuk Jang, PhD candidate. KAIST
View original imageExperimental results showed that LLMs trained through this process maintained high safety even when user data consisted entirely of dangerous questions and answers in extreme environments.
In particular, the rate at which the retrained model generated dangerous responses was 8 percent, which is even lower than the 18 percent observed in the original LLM that had not been retrained.
The research team expects that by achieving both customized performance and top-level safety without separate safety retraining or increased computational costs, the Buffer-and-Reinforce learning framework can be efficiently applied to personalized AI services.
Professor Kim stated, “The biggest strength of our learning framework is that when the LLM is temporarily set to jailbreak, it does not additionally learn dangerous questions or answers, while normal user data can be smoothly trained. This is expected to greatly contribute to creating a 'trustworthy AI service environment' required in the era of AI personalization and AI agents.”
Hot Picks Today
"Two Spend 500,000 Won in a Flash, Yet Wait Two Hours?" ... 'Dreamland' Shunned by Kids and Foreign Tourists
- The Era of Getting 200 Million Won After 30 Years Is Over... A Japanese Bank's "Perpetual Department Head" Who Just Clocked In Cries Out
- Spain Hit by Its Worst Tragedy... Lee Kangin, Moreno and Real Madrid Join Soccer World in Mourning
- They Said, “Make It Your Home for Life” - Then Suddenly Kicked 200 Passengers Off... What Happened to the Around-the-World Cruise?
- "130 Rice Bowls Sold Out in a Day"... 'Unfiltered Face' Adds 8 Million Won to Night Market Sales: China's Bungeunwoo
Meanwhile, Seokil Ham, a doctoral candidate at KAIST’s Department of Electrical and Electronic Engineering, is the first author of this research. The results have been selected as a Spotlight Paper at the International Conference on Machine Learning, a leading global conference in the AI field. The conference selected the top 2.2 percent of submitted papers this month for Spotlight status.
© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.