• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.34 no.3
• /
• pp.451-461
• /
• 2024

Recently, various AI chatbots based on large language models have been released. Chatbots have the advantage of providing users with quick and easy information through interactive prompts, making them useful in various fields such as question answering, writing, and programming. However, a vulnerability in chatbots called "prompt injection attacks" has been proposed. This attack involves injecting instructions into the chatbot to violate predefined guidelines. Such attacks can be critical as they may lead to the leakage of confidential information within large language models or trigger other malicious activities. However, the vulnerability of Korean prompts has not been adequately validated. Therefore, in this paper, we aim to generate malicious Korean prompts and perform attacks on the popular chatbot to analyze their feasibility. To achieve this, we propose a system that automatically generates malicious Korean prompts by analyzing existing prompt injection attacks. Specifically, we focus on generating malicious prompts that induce harmful expressions from large language models and validate their effectiveness in practice.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.228-233
• /
• 2023

인간의 수동 평가 시 시간과 비용의 소모, 주석자 간의 의견 불일치, 평가 결과의 품질 등 불가피한 한계가 발생한다. 본 논문에서는 맥락을 고려하고 긴 문장 입출력이 가능한 ChatGPT를 활용한 한국어 요약문 평가가 인간 평가를 대체하거나 보조하는 것이 가능한가에 대해 살펴보았다. 이를 위해 ChatGPT가 생성한 요약문에 정량적 평가와 정성적 평가를 진행하였으며 정량적 지표로 BERTScore, 정성적 지표로는 일관성, 관련성, 문법성, 유창성을 사용하였다. 평가 결과 ChatGPT4의 경우 인간 수동 평가를 보조할 수 있는 가능성이 있음을 확인하였다. ChatGPT가 영어 기반으로 학습된 모델임을 고려하여 오류 발견 성능을 검증하고자 한국어 오류 요약문으로 추가 평가를 진행하였다. 그 결과 ChatGPT3.5와 ChatGPT4의 오류 요약 평가 성능은 불안정하여 인간을 보조하기에는 아직 어려움이 있음을 확인하였다.

• Journal of KIBIM
• /
• v.13 no.3
• /
• pp.21-29
• /
• 2023

ChatGPT, a chatbot based on GPT large language models, has gained immense popularity among the general public as well as domain professionals. To assess its proficiency in specialized fields, ChatGPT was tested on mainstream exams like the bar exam and medical licensing tests. This study evaluated ChatGPT's ability to answer questions related to Building Information Modeling (BIM) by testing it on Korea's BIM expertise exam, focusing primarily on multiple-choice problems. Both GPT-3.5 and GPT-4 were tested by prompting them to provide the correct answers to three years' worth of exams, totaling 150 questions. The results showed that both versions passed the test with average scores of 68 and 85, respectively. GPT-4 performed particularly well in categories related to 'BIM software' and 'Smart Construction technology'. However, it did not fare well in 'BIM applications'. Both versions were more proficient with short-answer choices than with sentence-length answers. Additionally, GPT-4 struggled with questions related to BIM policies and regulations specific to the Korean industry. Such limitations might be addressed by using tools like LangChain, which allow for feeding domain-specific documents to customize ChatGPT's responses. These advancements are anticipated to enhance ChatGPT's utility as a virtual assistant for BIM education and modeling automation.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.24 no.1
• /
• pp.89-95
• /
• 2024

Nowadays, the competitiveness of a company depends on the ability of all organizational members to share and utilize the organizational knowledge accumulated by the organization. As if to prove this, the world is now focusing on ChetGPT service using generative AI technology based on LLM (Large Language Model). However, it is still difficult to apply the ChetGPT service to work because there are many hallucinogenic problems. To solve this problem, sLLM (Lightweight Large Language Model) technology is being proposed as an alternative. In order to construct sLLM, corporate data is essential. Corporate data is the organization's ERP data and the company's office document knowledge data preserved by the organization. ERP Data can be used by directly connecting to sLLM, but office documents are stored in file format and must be converted to data format to be used by connecting to sLLM. In addition, there are too many technical limitations to utilize office documents stored in file format as organizational knowledge information. This study proposes a method of storing office documents in DB format rather than file format, allowing companies to utilize already accumulated office documents as an organizational knowledge system, and providing office documents in data form to the company's SLLM. We aim to contribute to improving corporate competitiveness by combining AI technology.

• Journal of Intelligence and Information Systems
• /
• v.29 no.2
• /
• pp.219-240
• /
• 2023

Conversational AI which allows users to interact with satisfaction is a long-standing research topic. To develop conversational AI, it is necessary to build training data that reflects real conversations between people, but current Korean datasets are not in question-answer format or use honorifics, making it difficult for users to feel closeness. In this paper, we propose a conversation dataset (KOMUChat) consisting of 30,767 question-answer sentence pairs collected from online communities. The question-answer pairs were collected from post titles and first comments of love and relationship counsel boards used by men and women. In addition, we removed abuse records through automatic and manual cleansing to build high quality dataset. To verify the validity of KOMUChat, we compared and analyzed the result of generative language model learning KOMUChat and benchmark dataset. The results showed that our dataset outperformed the benchmark dataset in terms of answer appropriateness, user satisfaction, and fulfillment of conversational AI goals. The dataset is the largest open-source single turn text data presented so far and it has the significance of building a more friendly Korean dataset by reflecting the text styles of the online community.